UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

A human-inspired controller for robot-human object handovers : a study of grip and load forces in handovers… Chan, Wesley Patrick 2012

Your browser doesn't seem to have a PDF viewer, please download the PDF to view this item.

Item Metadata

Download

Media
24-ubc_2012_fall_chan_wesley.pdf [ 2.75MB ]
Metadata
JSON: 24-1.0073332.json
JSON-LD: 24-1.0073332-ld.json
RDF/XML (Pretty): 24-1.0073332-rdf.xml
RDF/JSON: 24-1.0073332-rdf.json
Turtle: 24-1.0073332-turtle.txt
N-Triples: 24-1.0073332-rdf-ntriples.txt
Original Record: 24-1.0073332-source.json
Full Text
24-1.0073332-fulltext.txt
Citation
24-1.0073332.ris

Full Text

A HUMAN-INSPIRED CONTROLLER FOR ROBOT-HUMAN OBJECT HANDOVERS A Study of Grip and Load Forces in Handovers and the Design and Implementation of a Novel Handover Controller by Wesley Patrick Chan B.A.Sc., The University of British Columbia, 2010 A THESIS SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF  MASTER OF APPLIED SCIENCE in The Faculty of Graduate Studies (Biomedical Engineering)  THE UNIVERSITY OF BRITISH COLUMBIA (Vancouver) October 2012  © Wesley Patrick Chan, 2012  Abstract Handing over objects is a common basic task that arises between people in many cooperative scenarios. On a daily basis, we effortlessly and successfully perform countless unscripted handovers without any explicit communication. However, handing over an object to a person is a challenging task for robotic “hands”, and the resulting interaction is often unnatural. To improve human-robot cooperation, the work described in this thesis has led to the design of a human-inspired handover controller based on analysis and characterization of the haptic interaction during human-to-human object handover. The first experiment in this thesis documents novel experimental work done to measure the dynamic interaction in human-human handovers. The grip forces and load forces experienced by the giver and the receiver during a handover are examined, and the key features are identified. Based on these experimental results, guidelines for designing humanrobot handovers are proposed. Next, this thesis describes a handover controller model that enables robots to hand over objects to people in a safe, efficient, and intuitive manner, and an implementation of the handover controller on a Willow Garage PR2 robot is documented. Finally, a second experiment is presented, which compares various tunings of the novel controller in a user study. Results show that the novel controller yields more efficient and more intuitive robot-to-human handovers when compared to existing handover controllers.  ii  Preface The work presented in this thesis was performed under the supervision of Dr. Elizabeth Croft and Dr. Machiel Van der Loos. Dr. Chris Parker provided guidance to the work documented in Chapter 3, Chapter 4, and Chapter 5. Dr. Amir Haddadi also provided advice to the work documented in Chapter 5. Parts of Chapter 3 have been published in the following publications: W. P. Chan, C. A. C. Parker, H. F. M. Van der Loos, and E. A. Croft, “Grip Forces and Load Forces in Handovers: Implications for Designing Human-Robot Handover Controllers,” in Proceedings of the 7th ACM/IEEE International Conference on Human-Robot Interaction, 2012, pp. 9-16. W. P. Chan, C. A. C. Parker, H. F. M. Van der Loos, and E. A. Croft, “Teaching Robots How to Share: Grip Forces and Load Forces in Handovers,” in Proceedings of the 2012 Human-Robot Interaction Pioneers Workshop, 2012, pp. 42-43. The author built the experimental apparatus, conducted the study, and performed the data analyses. Dr. Chris Parker assisted with the design and construction of a prototype of the apparatus. He also provided template code for the data acquisition program. The handover controller design presented in Chapter 4 is filed in the following provisional patent application with the author as the main contributor: E. A. Croft, W. P. Chan, C. A. C. Parker, and H. F. M. Van der Loos, “Method and System for Releasing an Object Held by a Robotic End Effector,” U.S. Patent Application 61/696,671, 2012 (Provisional). The human-subject studies documented in Chapter 3 and Chapter 5 were approved by the University of British Columbia Behavioural Research Ethics Board (H10-00503).  iii  Table of Contents Abstract................................................................................................. ii Preface..................................................................................................iii Table of Contents ................................................................................ iv List of Tables .....................................................................................viii List of Figures ..................................................................................... ix List of Symbols & Abbreviations ....................................................... x Acknowledgements ............................................................................ xii 1. 1.1. 1.2. 1.3. 1.4. 1.5.  2. 2.1. 2.2. 2.3. 2.4.  3.  Introduction ................................................................................. 1 Human-Robot Interaction .......................................................................... 2 Object Handovers ...................................................................................... 3 Research Goals .......................................................................................... 4 Contributions ............................................................................................. 5 Thesis Outline............................................................................................ 5  Background and Motivation ...................................................... 8 Robotics Literature on Handovers............................................................. 8 Existing Human-Robot Handover Controllers ........................................ 10 Human Tactile Sensing ........................................................................... 13 Haptics Literature on Object Manipulation............................................. 16  Human-Human Object Handover Study ................................ 18  3.1. Research Questions and Hypotheses ....................................................... 18 3.2. Method ..................................................................................................... 20 3.2.1. Apparatus .......................................................................................... 20 3.2.2. Force Sensing and Data Collection .................................................. 21 3.2.3. Volunteer Recruitment ..................................................................... 23 3.2.4. Experimental Setup and Procedure................................................... 23 iv  3.3. Data Analyses .......................................................................................... 26 3.4. Results ..................................................................................................... 28 3.4.1. Grip Force/Load Force Ratios .......................................................... 28 3.4.2. Grip Force versus Load Force .......................................................... 30 3.4.3. Transfer Time and Maximum Load Transfer Rate .......................... 32 3.4.4. Maximum Excess Receiver Load (MERL) ...................................... 34 3.5. Discussion................................................................................................ 35 3.5.1. Human Grip Force Control Strategy for Handovers ........................ 35 3.5.2. Transfer Time and Maximum Load Transfer Rate .......................... 37 3.5.3. Maximum Excess Receiver Load (MERL) ...................................... 38 3.5.4. Implications on Robot Handover Controllers................................... 38 3.6. Summary.................................................................................................. 39  4.  Robot Handover Controller Design......................................... 41  4.1. Handover Controller Design ................................................................... 41 4.1.1. Grip Force Control Function ............................................................ 42 4.1.2. Measured Load Force and Dynamic Effects .................................... 44 4.1.3. Handover Controller Flow Chart ...................................................... 46 4.2. Controller Parameters .............................................................................. 47 4.2.1. Zero Load Grip Force ....................................................................... 47 4.2.2. Release Force Threshold .................................................................. 48 4.2.3. Acceleration Tolerance ..................................................................... 49 4.3. Discussion................................................................................................ 49 4.4. Summary.................................................................................................. 50  5.  Robot-Human Object Handover Study .................................. 52  5.1. Handover Controller Implementation ..................................................... 53 5.1.1. Hardware Platform............................................................................ 53 5.1.2. Load Force Sensing .......................................................................... 54 5.1.3. Grip Force Control ............................................................................ 55 5.1.4. Acceleration Measurements ............................................................. 57 5.2. Controller Tunings .................................................................................. 58 5.3. Research Questions and Hypotheses ....................................................... 60 5.4. Method ..................................................................................................... 61 5.4.1. Experimental Design ........................................................................ 61 v  5.4.2. Apparatus and Data Collection ......................................................... 62 5.4.3. Survey Design ................................................................................... 62 5.4.4. Volunteer Recruitment ..................................................................... 63 5.4.5. Experimental Setup and Procedure................................................... 63 5.4.6. Data Analyses ................................................................................... 65 5.5. Results ..................................................................................................... 66 5.5.1. Survey Responses ............................................................................. 66 5.5.2. Grip Force and Load Force Data ...................................................... 68 5.5.3. Grip Force versus Load Force .......................................................... 70 5.5.4. Zero Load Grip Force and Maximum Excess Receiver Load (MERL)............................................................................................. 71 5.5.5. Transfer Time and Maximum Load Transfer Rate .......................... 72 5.6. Discussion................................................................................................ 73 5.6.1. Survey Responses ............................................................................. 73 5.6.2. Grip Force and Load Force Data ...................................................... 75 5.6.3. Maximum Excess Receiver Load (MERL) ...................................... 76 5.6.4. Transfer Time and Maximum Load Transfer Rate .......................... 78 5.6.5. Recommendations............................................................................. 79 5.7. Summary.................................................................................................. 81  6. 6.1. 6.2. 6.3. 6.4.  Conclusion .................................................................................. 82 Human-Human Object Handover............................................................ 82 Robot-Human Object Handover.............................................................. 83 Contributions ........................................................................................... 83 Future Work............................................................................................. 84  Bibliography ....................................................................................... 86 Appendix A. Personal Communications ........................................ 92 A1. Correspondence with Dr. Travis Deyle Regarding Their Handover Controller ................................................................................................. 92 A2. Correspondence with Dr. Andrea Mason Regarding Baton Size Design ...................................................................................................... 96  vi  Appendix B. Human-Human Object Handover Study Materials ..................................................................... 97 B1. Participant Consent Form ........................................................................ 97  Appendix C. Robot-Human Object Handover Study Materials .................................................................... 99 C1. Participant Consent Form ........................................................................ 99 C2. Experiment Survey ................................................................................ 101  Appendix D. Wilcoxon Signed-Rank Test Results for RobotHuman Handover Study Survey Responses ......... 103  vii  List of Tables Table 1. Force measurements extracted from baton data. ............................................... 22 Table 2. Measured transfer time, , maximum load transfer rate, ( ̇ ), and time to maximum load transfer rate, , for each condition. ..................... 33 Table 3. Analysis of variance (ANOVA) results for comparing transfer time, , maximum load transfer rate, ( ̇ ), and time to maximum load transfer rate, . ............................................................................................................ 33 Table 4. Tuneable parameters of the proposed handover controller. ............................. 47 Table 5. Coefficients for the robot grip force model. ........................................................ 56 Table 6. Analysis results for comparing likelihood of dropping the object. ................... 66 Table 7. Analysis results for comparing ease of taking the object. .................................. 66 Table 8. Analysis results for comparing user preference. ................................................. 66 Table 9. Measured zero load grip force, , release force threshold, , and maximum excess receiver load percentage, MERL%, for the four controllers. .............. 71 Table 10. Analysis results for comparing zero load grip force, , and maximum excess receiver load percentage, MERL%......................................................... 72 Table 11. Measured transfer time, , maximum load transfer rate, ( ̇ ), and time to maximum load transfer rate, , for each controller.............. 73 Table 12. Analysis results for comparing transfer time, , maximum load transfer rate, ( ̇ ), and time to maximum load transfer rate, . ... 73 Table D1. Analysis results for comparing likelihood of dropping the object. .............. 103 Table D2. Analysis results for comparing ease of taking the object. ............................. 103 Table D3. Analysis results for comparing user preference. ............................................ 103  viii  List of Figures Figure 1. Tactile afferents in the human hand. .................................................................. 14 Figure 2. A: Diagram of baton used in experiment. B: Side view of baton showing the FSRs for measuring grip forces. ....................................................................... 20 Figure 3. Table setup for the human-human handover experiment. ............................... 23 Figure 4. Giver grasping the baton on the bottom handle using a precision grip, with thumb placed over the FSR............................................................................... 24 Figure 5. The receiver load force, , load transfer rate, ̇ , giver grip force, , and receiver grip force, , from a single handover trial in the Heavy condition. ............................................................................................................................. 26 Figure 6. Load force, grip force, and grip/load ratio of giver and receiver for ten trials in the Heavy condition. ...................................................................................... 28 Figure 7. Grip force versus load force plots of one pair of giver (A) and receiver (B). . 30 Figure 8. Grip force versus load force plots for three different pairs of giver (A) and receiver (B) showing variations in the slope of the curves. ............................ 31 Figure 9. The average receiver load force, , load transfer rate, ̇ , giver grip force, , and receiver grip force, , in Light, Medium, and Heavy conditions. 32 Figure 10. Frequency distributions of MERL% in all conditions. .................................. 34 Figure 11. Proposed grip force control function. ............................................................... 42 Figure 12. Free body diagrams of the object in cases when A. acceleration is negligible, and B. acceleration is non-negligible. ............................................................... 44 Figure 13. Handover controller flow chart......................................................................... 46 Figure 14. Willow Garage PR2 robot used in the robot-human handover experiment. 53 Figure 15. Two-link model of robot forearm and gripper for calculating load force. ... 54 Figure 16. Four different tunings of the handover controller compared in the robothuman handover study ...................................................................................... 58 Figure 17. Table setup for the robot-human handover experiment. ............................... 63 Figure 18. The receiver load force, , load transfer rate, ̇ , giver grip force, , and receiver grip force, , for all controllers from one participant. ......... 69 Figure 19. Grip force versus load force plots for all controllers from one participant during object transfer. ....................................................................................... 70  ix  List of Symbols & Abbreviations Symbol  Description Object acceleration vector  α  Acceleration tolerance, significance level in statistical analyses  ANOVA  Analysis of variance Object acceleration in the vertical direction  DoF  Degrees of freedom Release force threshold  εcontact  Grip force threshold for identifying time of receiver contact  εrelease  Grip force threshold for identifying time of giver release Applied grip force Grip force control function Giver grip force Giver applied force vector on the object Giver load force Initial grip force Load force at end effector Initial load force Receiver load force Receiver grip force Receiver applied force vector on the object Receiver load force  ̇ FSR  Load transfer rate Force sensing resistor Zero load grip force Gravitational acceleration vector x  Symbol  Description Gravitational acceleration constant Grip force model of robot gripper  IAD  Intelligent assist device Stiffness coefficient of robot gripper Length from wrist joint to gripper center of mass Length from wrist joint to gripper tool center Slope of the grip force control function Object mass ( ̇ )  Maximum load transfer rate  MERL  Maximum excess receiver load  MERL%  Maximum excess receiver load as a percentage of the object weight Weight of the top half of the baton Robot wrist torque Robot wrist torque measurement error Robot wrist torque measurement offset Timestamp of receiver contact Time from receiver contact to maximum load transfer rate Timestamp of giver release Object transfer time Weight of robot gripper  xi  Acknowledgements I would like to thank my supervisors, Dr. Elizabeth Croft and Dr. Machiel Van der Loos, who have given me the freedom to explore and guided me with their wisdom throughout my Masters studies. They have provided me with continual support over the past two years, and to them, I offer my enduring gratitude. I would like to thank Dr. Chris Parker for his mentorship during my Masters studies. Dr. Chris Parker has inspired this thesis project and has devoted much time and effort to the students at the CARIS lab. My thanks also go to Dr. Amir Haddadi, who has provided invaluable advice to this thesis work. Many members of the CARIS lab have provided invaluable comments and feedback to help shape this thesis. I would like to express my gratitude to Benjamin Blumer, Navid Shirzad, Matthew Pan, Eric Pospisil, and AJung Moon. Many of these individuals have also provided help and support to various aspects of my thesis work. I would like to thank the individuals who have volunteered their time to participate in my studies and acknowledge the financial support provided by the Natural Sciences and Engineering Research Council of Canada, General Motors, and the Institute for Computing, Information and Cognitive Systems. I would also like to thank the University-Industry Liaison Office and Roger Miller for assisting us with the patent application process. Finally, I would like to thank my parents and grandparents for their continual, unconditional support.  xii  1. Introduction Traditionally, robots have been limited to working in highly structured environments, such as automated assembly lines, and designed to operate with minimal interaction with humans. Restricted to working in highly structured environments, these robots perform their tasks using pre-programmed trajectories. However, recently there has been a shift in this paradigm. Over the past decade, there has been increasing interest in the development of robots capable of operating outside of their confined spaces and working alongside people. The development of cooperative robots provides several benefits. For example, we can increase workplace productivity by leveraging the different capabilities of humans and robots. Humans are more skilled in performing dexterous manipulation tasks and they tend to be better decision makers. Robots, on the other hand, excel in high-precision and repetitive tasks; they can be better equipped to handle large, heavy objects, and are more suitable for carrying out dangerous jobs. For instance, certain operations within an automobile assembly line involve high variability in the assembly processes. These areas of the assembly line provide opportunities for assistive robots; in some operations, robots are used to perform heavy lifting tasks, while skilled workers perform installation and inspection of components that require dexterity and decision making capabilities beyond current machine abilities [1]. Cobots, a class of intelligent assist device (IAD), assist workers by allowing them to lift heavy objects with less force and providing virtual surfaces for guiding motion. The use of cobots reduces the risk of work related injuries occurring due to overexertion or repetitive motion [2]. Cooperative robots can also provide home support. Many parts of the world currently have an ageing population, and according to a report published by the United Nations, population ageing is a global issue which will intensify in the twenty-first century [3]. With an increased life expectancy and decreased birth rate, the relative population of older persons requiring additional homecare support grows. A report by McMullin et al. examining labour force ageing and skill shortages in Canada stated that nursing is one of the occupations that 1  risk labour shortages as existing workers retire and healthcare costs rise [4]. In response, many countries have invested in research on healthcare robots for providing assistance in hospitals and homecare robots for helping seniors or motor-impaired individuals at home [5– 7]. Service robots can reduce the workload on nursing staff by carrying out less critical tasks, such as routine delivery of medical devices, medicines, and documents in hospitals, thus allowing professionally trained nurses to better utilize their time [5]. Since the 1990s, researchers have developed many different cooperative robot platforms, aimed at a variety of applications, including intelligent assistants for supporting human workers in the factory [1], [8], museum tour guides to engage visitors and ask questions [9], and nursing robots for taking care of patients in the hospital or assisting older persons with their activities of daily living [10]. Many of the these developments have been and continue to be supported by industry partners [1], [8], [11], reflecting the growing demand for service robots from industry [12–14]. According to market research reports, the value of the global service robotics market is expected to grow significantly [15]. It can be expected that, in the near future, robots working with people will become more common and deployment of cooperative robots will become more widespread.  1.1. Human-Robot Interaction In most applications, robot assistants will work in human environments and interact directly with their users. In this context, such robots will have to manage many uncertainties introduced by unstructured surroundings. In addition, since human behaviour is generally fairly complex to model, cooperative robots will also have to cope with unpredictable user behaviours. Traditional robots are not designed to cope with high levels of uncertainty as they usually follow a pre-scripted sequence of actions. As a result, such robots are not suited for interacting and cooperating with people. For cooperative robots to work efficiently, they must allow users to interact with them naturally and smoothly. For example, often there are multiple ways to complete a task, and users may not act in a manner that is compatible with the robot’s specific program. Thus, designing smooth human-robot interaction is still a challenging problem. 2  Cooperative robots need to adapt to various situations and users in order to achieve smooth interactions with people. As an example, Wilcox et al. [8] considered a robotic assistant working at an aerospace manufacturing factory. The assembly of airplane spars is a task where mechanics can develop highly individualized workflow patterns. To allow mechanics to carry out their work naturally, mechanics should not be required to conform to the robotic assistant’s work pattern. In their analysis, Wilcox et al. note that it is important to develop adaptive robotic assistants capable of adjusting to the work styles of each individual. In any cooperative situation, whenever a worker is required to learn and remember the operation patterns of a robot, the cognitive load on the worker increases and it distracts them from the actual task at hand. Furthermore, if the behaviour of a robot is counterintuitive to the user, it may cause miscommunication or user frustration. In certain cases, miscommunication could lead to dangerous situations. In order to allow robots to cooperate effectively and safely with humans, it is imperative that we develop robots capable of natural interaction with their users.  1.2. Object Handovers Many different types of interaction are found in cooperative tasks – for example, sharing a common resource, giving instructions to an assistant, or lifting a heavy device together. One common interaction that occurs in many cooperative scenarios is object handover – i.e., the transfer of an object from one agent to another. Object handover is a task in physical human-robot interaction that will occur frequently for cooperative robots; examples include handing over a wrench to a mechanic, passing out information booklets to visitors at the museum, or bringing a bottle of medication to a patient. Humans perform countless handovers every day, and even without careful planning, they generally complete each handover safely and effortlessly without dropping the object. However, handing over objects in unscripted situations is still a challenging task for robots. The agents involved in a handover are the giver and the receiver. The giver is defined as the agent presenting the object, with the intention of transferring the object to the receiver. 3  The receiver is defined as the agent who is accepting the object, and intentionally taking responsibility for the object. Consider the mechanics of a typical handover, such as handing over a glass of water. During the handover, as the glass of water being is transferred, the giver and the receiver must synchronize the taking and the releasing of the object very carefully. If the glass of water is released too soon, it may end up being dropped, whereas if the glass of water is released too late, the receiver may have to pull very hard on it, and end up spilling the contents when the giver finally releases. A major difference between humans and robots that allows people to perform handovers much more efficiently and safely compared to robots is the difference in their tactile sensing capabilities. Human fingers contain many tactile sensors that allow them to detect various signals such as shear forces and deformations at points of contact with objects. The abilities to sense slip and distal object contacts contribute greatly toward humans’ object manipulation skills [16]. Most robot systems, on the other hand, at best only have normal-force sensors on their grippers. As a result, robots still cannot hand over objects to humans in a safe and efficient manner, and this remains a hindrance to achieving effective human-robot cooperation.  1.3. Research Goals The goal of this research is to enable robots to perform object handovers with humans safely without dropping the object, efficiently without requiring excessive force from the receiver, and intuitively in unscripted situations. Since humans regularly perform successful handovers with other people, a good first approach is to make robots emulate a person when participating in a handover. In order to achieve this, this work will first seek to better understand how humans hand over objects. As tactile sensing plays a crucial role in object manipulation [17], [18], the investigation on human handovers will focus on the haptic interaction. After characterizing the dynamic interaction in human handovers, this thesis will then propose a novel robot handover controller design for achieving smooth handovers based on the experimental findings. Following that, the proposed controller will be implemented on a Willow Garage PR2 robot to validate the design. Finally, this thesis will present the results 4  of a user study conducted to compare how various tunings of the controller affect the resulting handovers.  1.4. Contributions This thesis presents two main contributions. Firstly, through the study of haptic interaction in human-human handovers, this work offers a better understanding of the dynamic interaction in handovers. Even though object handover is a basic, essential interaction, the dynamic interaction during a handover is a topic that has still not been wellinvestigated. This thesis provides a better understanding of the force interaction between the giver and the receiver during a handover. This knowledge can be used in robotics for designing smooth human-robot handovers. Furthermore, these findings can also potentially be used in medical research for a better understanding of the effects of sensorimotor diseases on dyadic object manipulation, and toward the treatment or rehabilitation of individuals with sensorimotor impairments. Secondly, this thesis presents a novel controller that enables robots to hand over objects to humans in a safe, efficient, and intuitive manner. This controller offers several advantages: it is hardware independent, does not require special sensors, and can be readily implemented onto most existing robot systems without any hardware modifications. An implementation of the controller on a PR2 robot is documented and a user study validates that the novel controller does produce safer and more natural handovers. With this controller, robots will be able to smoothly hand over objects to people, and will thus be able to work with people more effectively.  1.5. Thesis Outline In the remainder of this thesis, Chapter 2 provides a review of the literature related to handovers. Chapter 3 presents a user study on human-human handovers. Chapter 4 documents the design of a novel handover controller, followed by Chapter 5, which discusses 5  an implementation of the controller and a user study conducted to validate the controller. Finally, Chapter 6 provides a summary and conclusion to the work presented in this thesis. A brief summary of each chapter is provided below. The literature review in Chapter 2 encompasses works in robotics related to handovers. Existing studies focus on the kinematics of handovers and how a robot should approach the human and orient itself for a handover. Chapter 2 draws attention to the small number of handover controllers found in the literature and discusses the limitations of these controllers. As tactile sensing plays a crucial role in enabling humans to perform skilled handovers, this chapter also provides a review of the haptics literature related to object manipulation. Most studies found in the haptics literature focus on single-person object manipulation tasks – e.g., how people control their grip force when picking-up an object and placing it down. It has been shown that, during pick-and-place tasks, people tune their grip force according to the experienced load force. However, the interaction in a handover, which involves two individuals, is still not well-studied. To design smooth human-robot handovers, we first need to better understand the dynamic interaction in human-human handovers. To address the lack of knowledge in handover interaction, Chapter 3 presents a user study on human-human handovers. This study investigates and characterizes the haptic interaction in handovers. The hypothesis is that during a handover, the giver and receiver employ a grip force control strategy similar to the one used in pick-and-place tasks. The results demonstrate that, although the strategies used for a pick-and-place task and a handover are similar, there are some important differences. Key features of the interaction are identified and cues used for coordinating object transfer are discussed. The experimental findings then provide the basis for several guidelines on designing safe and natural humanrobot handovers. Using the results from Chapter 3, Chapter 4 documents the design of a novel handover controller for enabling robots to hand over objects to people in a safe, efficient, and intuitive manner. The controller exhibits human behaviour and allows the robot to emulate a person during a handover. A couple of tuneable parameters allow the controller to be adjusted to suit different situations and properties of the object being handed over. This chapter discusses the implications of each parameter on the resulting handovers. 6  To validate the novel controller proposed in Chapter 4, the controller is implemented and tested on a Willow Garage PR2 robot. Chapter 5 documents the implementation as well as the user study conducted to compare different tunings of the controller. The study tests the hypothesis that users prefer a more human-like controller tuning. Chapter 5 discusses that, although certain tunings yielding more human-like handovers are preferred by people, there are important trade-offs between safety of the object and ease of taking by the receiver. To conclude this thesis, Chapter 6 provides a summary of the presented work, discusses the limitations of the controller, and outlines possible future work.  7  2. Background and Motivation This chapter commences with a review of the robotics literature related to handovers in Section 2.1 and describes existing handover controllers in Section 2.2. The latter discusses the limitations of these controllers and identifies the need for an improved design based on performance criteria. In order to design smooth human-robot handovers, the first step is to understand how humans perform handovers, and, in general, how they manipulate objects. Since humans rely heavily on tactile information for object manipulation, Section 2.3 provides an overview of human tactile sensing capabilities, followed by a review of the haptics literature related to object manipulation in Section 2.4.  2.1. Robotics Literature on Handovers To date, studies on human-robot handovers have focused on how a robot should approach a person, and how it should reach towards a person to hand over an object [19–26]. Individual groups of researchers studying handover reaching motions have shown that people prefer working with robots that employ human-like trajectories [19], [20]. In a study by Shibata et al., researchers compared different velocity patterns using a one degree of freedom (DoF) robot [19]. Their experimental setup involved a linear robot on a table handing over a glass to a human receiver. The researchers compared bell-shaped velocity patterns, which model the minimum jerk trajectory of human reaching motions, to triangular and trapezoidal velocity patterns representing typical robot trajectories. Results showed that when the robot used a human-like bell-shaped velocity pattern, people found it easier to receive the object, and they described the robot as more pleasant, more careful, and more skilled. Extending Shibata et al.’s work, Huber et al. examined the reaching motion during a handover in three dimensions and found similar results [20]. In Huber et al.’s study, the experimenters compared a trapezoidal joint velocity profile with a Cartesian minimum jerk trajectory. Results indicated that when the robot used a minimum jerk trajectory, receivers 8  felt safer during the experiment, and had shorter reaction times for taking the object. The studies by Shibata et al. and Huber et al. concluded that the application of human-like motions to robots can be an effective strategy for human-robot handovers. Examining a different aspect of handovers, Cakmak et al. studied human preference for robot configuration during handovers [23]. Their study explored how a robot should position itself for handovers and how it should hold the object. The researchers compared configurations that were learned from human examples against configurations that were planned using a human kinematic model. Their results showed that although planned configurations provided better reachability of the object, people preferred learned configurations, and they felt that learned configurations appeared more natural and more appropriate. Cakmak et al. explained that participants’ preferences were given in consideration of the object’s function, and the researchers illustrated this point with a specific example: when using the planned configuration, a mug was handed over facing down. The configuration chosen for the mug was based on the receiver’s comfort of grasping, assuming that the mug was to be carried in the same orientation. However, people often grasped the mug in a different way such that it could be rotated upright. The experimenters concluded that although a human kinematic model-based approach produces practical configurations, these configurations are lacking in terms of usability, appropriateness, and naturalness. In a different study by Cakmak et al. exploring the use of robot postures and gestures in handovers, they found that certain postures can better communicate a robot’s intent [24]. The researchers conducted an online survey asking participants to classify different postures of a robot holding an object as carrying, handing over, looking at, or showing the object. Results showed that postures with an extended arm were most frequently classified as handing over. Following the survey, Cakmak et al. conducted a user study to examine the use of gestures in handovers. In the experiment, the researchers varied the postures used by the robot for carrying and handing over an object. The experimenters reported that when postures labeled as carrying in the survey were used for carrying, and postures labeled as handing over were used for handing over, the robot’s intention to hand over the object was best conveyed. In these cases, the robot’s reaching gesture had a larger movement, and the amount of time the receiver waited before taking the object was shorter. 9  Dehais et al. studied the use of a human-aware trajectory planner [27] for planning the reaching motion of a handover [28]. The human-aware trajectory planner evaluates the legibility, safety, and physical comfort of robot motions taking into account a person’s position and orientation. In an experiment, the researchers compared trajectories generated by the human-aware planner with straight-line trajectories. Results indicated that motions generated using the planner required less physical effort from receivers, and were rated as more legible and safer. In summary, Dehais et al.’s study demonstrated the importance of taking the human into consideration when planning for human-robot handovers. Koay et al. also studied how a robot should approach a person when preparing to handover an object [22]. In their experiment, a robot was programmed to approach the person receiving the object from different angles. Koay et al. reported that most participants preferred the robot to approach them from the front and hand over the object from the front since it provides the most visibility of the robot’s motion. To summarize, existing studies on human-robot handover have shown that people prefer working with human-like robots. When a robot mimics human gestures and postures during a handover, people feel safer and more natural working with the robot. Furthermore, these studies have also indicated the importance of taking the object’s function and receiver’s state into account. It can be concluded that, when approaching a person and reaching over to hand over an object, a robot should use human-like motions and configurations.  2.2. Existing Human-Robot Handover Controllers While the kinematics of the reaching motion prior to object transfer has been well studied, the dynamic interaction during the transfer of the object has still not been considered; we still do not know how a robot should approach object transfer. A few robot controllers for transferring objects can be found in the literature; however, most of these studies have not focused on object transfer and, therefore, these controllers often have many limitations. This section provides a review of existing controllers for transferring objects, and identifies the need for a more robust handover controller.  10  A variety of approaches to object transfer can be found in the literature [7], [29–33]. In a study investigating people’s responses to robot gesture during handovers, Edsinger and Kemp used a time-based handover controller [29]. Their study involved a robot taking a box from the participants and handing the box back. The objective was to determine whether people will interpret the robot’s reaching motion as a cue to take the object. When handing over the box, in one case, the robot would reach above the table, say “Done”, wait for one second, and then release the box. In another case, the robot would perform an identical action, but would reach just past the table edge. Edsinger and Kemp reported that when giving the object to the robot, participants automatically aligned the box with the robot’s hand, making it easier for the robot to take the object. However, when the box was handed back to the participants, while participants only allowed the robot to drop the box in 1 out of 30 trials when the robot reached past the table, they allowed the robot to drop the object onto the table 11 out of 30 trials. Edsinger and Kemp’s study showed that in certain situations a time-based handover controller can results in a high rates of dropping the object. Nagata et al. developed a grasping algorithm for receiving objects from, and handing over objects to humans [30]. Their platform consisted of a four-fingered gripper with individual 6-axis force-torque sensors at each fingertip. Using force-torque data from each fingertip, the grasping algorithm continually evaluates the grasp stability by monitoring finger slippage. The algorithm takes into account manipulability, joint angle range, and contact surface geometry. During grasping, if the stability constraint of a finger is violated, the algorithm repositions the finger to maintain a stable grasp. Using this grasping algorithm, the robot receives an object when a person pushes the object into the fingers. Upon sensing contact at one of the fingers, the robot softly closes its fingers, as the person adjusts the object into a comfortable position and orientation in the robot’s hand. The robot then begins to monitor grasp stability and reposition each finger until a stable grasp is achieved. To take the object from the robot, the person is required to pull on the object such that it becomes unstable in the robot’s hand. The robot then lets go of the object when stability conditions are no longer satisfied. Although this algorithm allows the robot to continually maintain grasp stability, the implementation requires individual 6-axis force torque sensors, which can be quite costly, for each fingertip. Furthermore, the receiver must forcibly take the object out of the robot’s grippers; this decreases the efficiency and smoothness of the handover since the 11  receiver has to apply more force than necessary to take an object. Finally, any interaction that causes the object to be unstable in the robot’s hand will cause the robot to let go, regardless of whether this interaction is handover-related. To allow smoother handovers, Deyle et al. used a force/torque threshold-based approach for handing over objects [31]. In their study investigating the use of radio frequency identification (RFID) tags for guiding robots, the robot located tagged objects and delivered objects to receivers wearing tagged wristbands. According to Dr. Deyle (refer to Appendix A1), the robot used in the study had individual 6-axis force-torque sensors at the base of each finger. When handing over an object, the robot would release once the force/torque values exceeded a pre-set threshold. This approach has the advantages of limiting the amount of force required from the receiver to take the object and ensuring smoother and more efficient object transfers. However, Dr. Deyle noted that this method requires manual tuning of the threshold values and these values are dependent on the hardware configuration (e.g., longer fingers result in higher torques). This approach would also be susceptible to dropping the object due to collisions; for example, if the receiver accidentally bumped the object when reaching for it. To ensure object safety, Bohren et al. implemented a displacement-based method for their autonomous robot butler [32]. In their study, the robot was programmed to serve beverages to people. When handing over a drink, to avoid accidentally dropping it, the robot was programmed to release when a person pulls hard enough on the bottle such that the robot’s impedance-controlled arm displaces more than one centimeter vertically. Although this implementation prevents the object from accidentally dropping, it requires an excessive amount of force from the receiver to take the object. From a video demonstration [34], it can be seen that since the receiver needs to pull very hard on the object, the resulting jerk at the point of release is very high, and this causes carbonated drinks to spill out of their bottles when opened by the receiver. As seen from the examples presented above, there have been a variety of approaches to human-robot handovers, including time-based [29], [33], slip-based [30], force/torque threshold-based [7], [31], and displacement threshold-based [32] controllers. Existing controllers either emphasize object safety or handover smoothness, and trade-off between the 12  two criteria. As a result, human-robot handovers are not yet as safe and efficient as humanhuman handovers. To allow effective human-robot cooperation, a robot handover controller that can produce safe, efficient, and intuitive handovers is still needed.  2.3. Human Tactile Sensing During a handover, as the object is being transferred from a giver to a receiver, the two individuals must synchronize the taking and releasing of the object carefully in order to ensure a safe and smooth object transfer. To achieve the precise timing required, the giver and the receiver must communicate with each other, either explicitly or implicitly. Humans often perform successful handovers without verbal communications, and can do so, even when blindfolded, relying only on tactile feedback from the hand. Studies have shown that when people lose their sense of touch or proprioception, they experience great difficulties with simple object manipulation tasks, such as picking up an object from a table, and they become unable to regulate their grip forces appropriately [17], [18], [35]. Therefore, it can be inferred that haptic communication plays an important role in handovers.  13  Humans’ tactile sensing capabilities come from the large amount of tactile afferents (i.e., touch sensors) covering the inside of their hands [16]. Four functionally distinct types of tactile afferents have been identified in the human hand – FA-I (fast-adapting type I), SA-I (slowly-adapting type I), FA-II (fast-adapting type II), and SA-II (slowly-adapting type II). Figure 1 shows the different tactile afferent types found in the human hand along with their properties and density distributions.  Figure 1. Tactile afferents in the human hand. (Adapted from [16]). 14  FA-I and SA-I afferents terminate in superficial skin surfaces, and respond to skin deformations. FA-I is more sensitive to relatively high frequency deformations, while SA-I is more sensitive to lower frequency changes and can detect sustained deformations. Both FA-I and SA-I have a higher density at the fingertips, allowing the fingertips to easily extract spatial features such as the skin forming and breaking contact with objects, or detecting surface textures. FA-II and SA-II terminate deeper in the skin at the dermal and subdermal fibrous tissues. FA-II is specialized for sensing transient vibrations and can detect hand-held objects making contact or breaking contact with other objects. SA-II responds to remotely applied stretching of the skin, and can detect tangential shear in the skin that occurs during object manipulation. Both FA-II and SA-II have a low density compared to FA-I and SA-I and are more uniformly distributed. The different types of tactile afferents, together with the deformable skin, allow the human hand to detect various “events”, such as slippage and object contact, accurately in order to perform skilled object manipulation [16]. The human hand is capable of detecting localized slips (micro-slips) that occur just before an object begins to slip in the hand [17], [36], allowing people to safely and efficiently perform object manipulation tasks, including handovers. By comparison, there is a huge disparity between the tactile sensing capabilities of humans and that of current robots. Although various types of tactile sensors for robots have been developed, the scaling of the sensor size, and the processing of the sensor data still remains challenging [37]. The spatial resolution achieved by current robot tactile sensors are in the range of 10 sensors per cm2 [38] while human fingertips can have more than 100 afferents per cm2 [16]. Currently, the spatial resolution of robot tactile sensors is limited by the physical size of the sensor elements as well as the amount of wire needed to connect to each sensor. Increasing sensor density also increases the amount of data to be processed, limiting the speed at which information can be extracted. Furthermore, typical robot slip sensors operate at approximately 40 Hz [39] whereas human tactile afferents can detect deformations at frequencies up to 400 Hz [16], imposing yet another limitation. As a result, robots still cannot achieve human-level performance in tactile sensing, requiring robots to 15  utilize an alternate strategy for slip detection in object manipulation. The next section introduces grip force control strategies used by humans, and reviews the related work on object manipulation.  2.4. Haptics Literature on Object Manipulation The ability to accurately detect object slip enables humans to efficiently regulate their grip force when manipulating objects. Studies have shown that during object manipulation, the amount of applied grip force is tightly coupled to the experienced load force [17], [40– 44]. While robotic tactile sensors still cannot achieve the performance levels of human tactile afferents, most robotic arms and grippers are equipped with load force and grip force sensing capabilities [45–48]. If we can understand the coupling between grip force and load force during handovers, we may be able to program a robot to hand over objects based on detection of load force rather than slip. To understand how humans control grip force during object manipulation, this section provides a review of the related haptics literature. Studies have shown that people are capable of controlling their grip force appropriately and efficiently according to the amount of experienced load force during object manipulation [17], [40–44]. For example, Johansson and Westling studied how people control grip force when they pick up an object, hold on to it, and place it back on a table [17]. The experimenters found that during such pick-and-place tasks, people gradually increase their grip force as they take on the load of the object; the grip force is regulated according to the experienced load force such that a minimum amount required to prevent object slip is applied at each instance with a small additional amount of safety margin. Furthermore, Johansson and Westling showed that when there is a slow or sudden change in load force due to addition of mass or external forces on the object, people can adjust their grip force accordingly to maintain a roughly constant grip-force-to-load-force-ratio [17], [43]. Similarly, Flanagan et al. studied grip force control when moving grasped objects, and they also observed a tight coupling between grip force and load force [40]. The researchers found that when transporting an object, participants finely regulated the grip force in phase with their arm movement. Load force variations due to acceleration were accounted for to 16  prevent object slip. Based on their results, Flanagan et al. concluded that programming of grip force control is part of the planning process of arm movement for moving an object. Expanding on Flanagan et al.’s findings, Hermsdörfer et al. studied the effects of microgravity and hypergravity on grip force control [41]. Hermsdörfer et al. found that people are able to adjust their grip force to account for dynamic load forces under microgravity and hypergravity, even though the pattern of load force changes were completely different from those under normal gravity. Furthermore the coupling between grip force and load force was also preserved during transitions between gravity levels. Hermsdörfer et al.’s study showed that humans’ grip force control remains consistent even with massive environmental condition changes. A study by Mason and Mackenzie investigating grip force control during handovers also revealed strong coupling between grip force and load force [44]. In their experiment, subjects handed over an object of known weight with an accelerometer attached for measuring dynamic loads. Results showed that prior to object transfer, as the giver was reaching towards the receiver, the giver maintained a fairly constant grip-force-to-load-force-ratio. Similarly, after object transfer when the receiver brought back the object, the receiver also maintained a fairly constant grip-force-to-load-force-ratio. These results show that humans use a similar grip force control strategy for both point-to-point object transportation and for reaching over during a handover. However, since their apparatus did not have the capability of determining the individual load forces when both giver and receiver were simultaneously in contact with the object, the grip/load ratios during object transfer remained unknown. In summary, researchers have shown that the coupling between grip force and load force during object manipulation is consistent across many different situations. However, existing studies on grip force control have been limited to individual manipulation of an object, and we still do not have a good understanding of grip force control and the dynamic interaction during the object transfer phase of a handover. To address this knowledge gap, the next chapter presents an experiment that investigates and characterizes the haptic interaction in human-human handovers.  17  3. Human-Human Object Handover Study This chapter documents a user study conducted to investigate and characterize the haptic interaction in human-human object handovers1. An instrumented baton, designed for this experiment, recorded the grip and load forces of eighteen participants working in pairs to hand a weighted object back and forth. Data show that during object transfer, both the giver’s and the receiver’s grip forces are strongly coupled to their load forces. Furthermore, the data points toward an implicit social contract in which the giver is responsible for object safety and the receiver for maintaining handover efficiency. Results from this chapter provide a basis for the robot handover controller design described in Chapter 4. In the following, Section 3.1 gives a formal statement of the research questions and hypotheses. Section 3.2 discusses the experimental method. Section 3.3 documents the data analysis procedure, followed by Section 3.4, which presents the results. Finally, Section 3.5 provides a discussion of the findings, and Section 3.6 sums up this chapter.  3.1. Research Questions and Hypotheses The purpose of this study is to better understand the dynamic interaction in humanhuman handovers. The objective is to characterize the haptic interaction in successful  1  The study and results presented in this chapter have been published in:  W. P. Chan, C. A. C. Parker, H. F. M. Van der Loos, and E. A. Croft, “Grip Forces and Load Forces in Handovers: Implications for Designing Human-Robot Handover Controllers,” in Proceedings of the 7th ACM/IEEE International Conference on Human-Robot Interaction, 2012, pp. 9-16. W. P. Chan, C. A. C. Parker, H. F. M. Van der Loos, and E. A. Croft, “Teaching Robots How to Share: Grip Forces and Load Forces in Handovers,” in Proceedings of the 2012 Human-Robot Interaction Pioneers Workshop, 2012, pp. 42-43.  18  human-human handovers, and in particular, this study seeks to answer the following questions: 1. How do human givers and receivers regulate their grip forces with respect to their load forces during object transfer? 2. How does a human giver determine when it is safe to release the object during a handover task? 3. What are the consistent features in the giver’s and receiver’s grip force control strategies? To answer these questions, the experimenter conducted a study to examine the grip forces and load forces of the giver and the receiver during a handover. Existing studies have shown that when holding an object, humans maintain a fairly stable grip-force-to-load-force-ratio such that the minimum amount of grip force required to prevent slip is applied with a small additional safety margin [17], [49]. Furthermore, in the presence of load force changes, humans actively regulate their grip force to maintain the constant grip-load ratio [40], [42–44]. Based on these studies, the experimenter formulated the following hypotheses (superscript H denotes human-human handover study): : The  giver maintains an approximately constant grip-force-to-load-force-ratio during  object transfer. :  The receiver maintains an approximately constant grip-force-to-load-force-ratio during object transfer.  If the giver and receiver maintain stable grip-force-to-load-force-ratios during object transfer, it would imply that humans employ similar grip force control strategies for both holding and handing over objects, and that we can program robots to simply regulate their grip force proportional to the load force when handing over objects to people. Otherwise, a more complex interaction exists between the giver and the receiver during object transfer, and a more sophisticated robot handover controller is required.  19  3.2. Method Eighteen healthy adults (ten males and eight females) participated in a handover study (refer to Appendix B for consent form). Participants worked in pairs to hand over a baton to each other, and each pair of participants performed 120 handovers in total. An instrumented baton measured the giver’s and the receiver’s grip forces and load forces while the weight of the baton was varied as the independent variable.  3.2.1. Apparatus To measure the giver’s and receiver’s individual load forces during object transfer, an ATI (Apex, NC, USA) Mini45 6-axis force/torque sensor was mounted between two sheetaluminum boxes to create a baton for the experiment (Figure 2A). In the experiment, the giver and receiver grasped onto the aluminum boxes, while the ATI sensor measured the load forces of the two individuals. To measure the giver’s and receiver’s grip forces, an Interlink  Figure 2. A: Diagram of baton used in experiment. B: Side view of baton showing the FSRs for measuring grip forces. (Figure published in Chan et al. 2012 [50].) 20  Electronics (Camarillo, CA, USA) FSR 406 force sensing resistor (FSR) was attached to the side of each handle, with a thin plastic square placed over each FSR to distribute the pressure (see Figure 2B). The FSRs on the handles are located where the giver’s and receiver’s thumbs would naturally land when grasping the baton. To provide a consistent gripping surface on the handles, new pieces of cardstock were placed over each gripping surface for every pair of participants. To track the handover motion, Xsens (Culver City, CA, USA) MTx acceleration sensors were attached to the baton and the giver’s and receiver’s wrists. The experimenter varied the weight of the baton among three different conditions – Light (483 ± 1 g), Medium (577 ± 1 g), and Heavy (678 ± 1 g) – by attaching additional masses at the bottom of the baton. The size and weight of the baton were designed to resemble common everyday objects such as mugs and bottles. The dimensions of the handles (see Figure 2A) were chosen in consultation with Dr. Andrea Mason to allow ease of grasping (refer to Appendix A2).  3.2.2. Force Sensing and Data Collection The ATI Mini45 force/torque sensor has a force resolution of 0.125 N in all three axes [51]. Signals from the FSRs were first linearized using a four-stage amplifier circuit before being input to a Quanser (Markham, ON, Canada) Q8 board through two parallel 14bit analog-to-digital converter inputs. The FSRs were calibrated using the ATI force/torque sensor according to the procedure described in [52] after preloading them with a shear force of 9.81 N and a compressive force greater than 20 N for over 20 hrs. After calibration the FSRs were tested against the ATI force/torque sensor with shear forces varying up to 9.81 N, and the root-mean-square-errors were measured to be 1.5 N. The Xsens MTx accelerometers had a measured root-mean-square noise level of ~0.2 m/s2 and were connected to the data acquisition computer via an Xbus Master. During the experiment, a Quanser WinCon real-time kernel was used to record data from the Q8 board, and FSR data read from the Q8 board were written to shared memory on  21  the data acquisition computer. At the same time, a separate, synchronized process on the same computer read the ATI force/torque data, Xsens acceleration data, as well as the shared memory values, and recorded them to disk. All data were captured at 50 Hz. From the recorded data, the experimenter extracted the giver’s and receiver’s grip forces and load forces, as well as the load transfer rate. Table 1 shows a list of the extracted measurements. Table 1. Force measurements extracted from baton data. Variable  Description Giver load force Receiver load force  ̇  Load transfer rate Giver grip force Receiver grip force  In this study, the experimenter modelled the object transfer phase as a quasi-static process – i.e., dynamic forces in the upward direction due to accelerations were small enough to be negligible compared to the baton’s weight2. Since the baton was carried only by the giver and/or receiver at all times, the weight of the baton must equal the sum of the giver’s and receiver’s load forces. Thus, we can write: (1) where calculate compute  is the weight of the baton. As the baton was carried upright at all times, we can and  from the force/torque sensor’s z-axis force measurements,  , an offset was subtracted from  . To  : (2)  where the offset,  , reflects the weight of the top half of the baton, and was obtained  by averaging the  reading over the first 500 ms of the handover trial, when the baton was  2  Analysis of accelerometer data confirmed that average RMS value of baton acceleration during object transfer was less than 0.4 m/s2  22  held motionless by the giver. Rearranging Equation (1), giver load force can then be computed as: (3) The load transfer rate, ̇ , was computed by differentiating  with respect to time.  3.2.3. Volunteer Recruitment Participants for this study were recruited through advertisements posted around the University of British Columbia campus as well as through personal contacts. Seven of the nine pairs of participants were acquainted with each other.  3.2.4. Experimental Setup and Procedure The giver and receiver were seated across a table for the experiment. Figure 3 shows the table setup. Two square markers indicated the starting positions of the giver and receiver.  Figure 3. Table setup for the human-human handover experiment. Two square markers indicate the starting positions for the giver and receiver. An interception zone is marked off in the middle for object transfer. (Figure published in Chan et al. 2012 [50].) 23  A pair of parallel lines in the middle of the table marked an “interception zone”. At the start of each handover, the giver grasped the baton with the right hand on the lower handle, and held the baton above the nearest square marker. At the same time, the receiver’s right hand started from the other square marker. On a verbal “go” signal, the giver reached over to the middle of the table with the baton. Once the giver had reached the interception zone, the receiver then took the baton by grasping the top handle with the right hand. After object transfer, both giver and receiver then returned their hands to their starting positions to complete the handover. Participants were instructed to hold the baton with a precision grip using all their digits so that only the fingertips (intermediate and distal phalanges) were in contact with the baton, and to position their thumbs over the FSRs. Figure 4 shows a precision grip. This type of grip was used in this experiment for the following reasons: (i)  In the most relevant study on grip forces in handovers by Mason and Mackenzie [44], a precision grip was used. The current study expands on their findings.  Figure 4. Giver grasping the baton on the bottom handle using a precision grip, with thumb placed over the FSR. (Figure published in Chan et al. 2012 [50].) 24  (ii)  According to Zheng et al. [53], approximately 50% of all grips in toolintensive activities are precision grips.  (iii)  A precision grip allows grip forces to be measured more easily and accurately using only one sensor since there are only two opposing grasp surfaces.  In the experiment, participants were also asked to maintain a vertical orientation of the baton and avoid contacting the table. The grasp configuration was chosen to resemble handovers with various everyday objects such as cans, bottles, or drinking glasses. The speed of the handover was, however, not specified. For each pair of participants, one person was arbitrarily chosen to be the giver first and the other the receiver. Each giver performed two sets of ten handovers for each of the three weight conditions in randomized block trials. After a total of sixty handovers, the giver and the receiver switched roles and repeated another sixty handovers. To avoid fatigue, participants were given a rest period after every set of ten handovers and were allowed to take any additional breaks if needed. However, none required additional breaks.  25  3.3. Data Analyses Figure 5 shows the measurements extracted from a single handover trial in the Heavy condition. A fourth order low-pass Butterworth filter with a cut-off frequency of 14 Hz was used to filter all grip force data to reduce noise prior to analysis (a common procedure from the literature, e.g., [54–56]). From Figure 5, we can see that in the beginning of the handover when the giver was holding onto the baton, zero as expected. Note that  started high, while  and  started from  is equal to the object weight subtracted by  – see  Equation (3). During object transfer, as the load was being transferred from the giver to the receiver,  decreased, while  zero, and  rose to a stable value.  and  To answer the hypotheses, force-ratios were examined, i.e.,  increased. After object transfer,  and /  and  dropped to  , the giver’s and receiver’s grip-force-to-load/  . In addition, to gain further insight into  Figure 5. The receiver load force, , load transfer rate, ̇ , giver grip force, , and receiver grip force, , from a single handover trial in the Heavy condition. Transfer time is measured from to . (Figure published in Chan et al. 2012 [50].) 26  how the giver and receiver control their grip forces with respect to load forces during object transfer, the experimenter also examined the giver’s and receiver’s grip force versus load force plots. For each handover, to quantify the haptic interaction between the giver and the receiver, the experimenter calculated the transfer time, rate,  , the maximum load transfer  ( ̇ ), and the time from receiver contact to maximum load transfer rate,  Figure 5). Object transfer time,  , was calculated by finding the difference between  the time at which the receiver first made contact with the baton, the giver finally released the baton, contact time,  (see  , and the time when  (i.e.,  ). Receiver  , was determined by finding first instance when the receiver’s grip force  rose above a threshold εcontact, starting from the end of each handover data. Likewise, giver release time,  , was determined by finding the first instance when the giver’s grip force  dropped below a threshold εrelease. The values of εcontact and εrelease were chosen to be 0.5 N through inspection of the data to be as close to zero as possible while remaining above the noise level. The maximum load transfer rate, measured during object transfer, and the time when  ( ̇ ), was defined as the maximum ̇  was calculated as the time between  and  ( ̇ ) occurred.  Through observing the experimental data, it appeared that often, toward the end of object transfer, the maximum measured  was slightly greater than the total weight of the  object. Referring to Figure 5, we can see that the maximum measured object transfer is slightly greater than the  toward the end of  after object transfer. This excess amount of  load force may be a result of the grip force safety margin employed by the giver in the grip/load ratio, and that the giver is delaying the release of the object pass the point where has reached zero. To further investigate this, the experimenter also measured the maximum excess receiver load as a percentage of the object weight (MERL%). To identify any significant differences across conditions, the experimenter computed one-way repeated measures analyses of variance (ANOVA) for the dependent measures, with post hoc analyses using the Bonferroni correction (α = 0.05). The first set of ten handovers in each condition was treated as a training set and only the second set of ten handovers was 27  used for data analysis to mitigate any learning effects. However, examining the data from the first ten handovers showed no observable trends, suggesting that neither learning nor fatigue was a significant factor in the experiment.  3.4. Results 3.4.1. Grip Force/Load Force Ratios In Figure 6, ten representative handover trials in the Heavy condition for one giverreceiver pair are shown, with individual plots for giver load, grip/load ratio,  /  , receiver load,  , receiver grip,  , giver grip,  , giver  , and receiver grip/load ratio,  Figure 6. Load force, grip force, and grip/load ratio of giver and receiver for ten trials in the Heavy condition. Giver grip/load ratio, / , was fairly constant throughout the handover, but rose quickly toward the end of object transfer. Receiver grip/load ratio, / , started at a lower value in the beginning of object transfer and rose to a fairly stable value. Dashed lines show the transfer period from one trial. (Figure published in Chan et al. 2012 [50].) 28  /  . The  /  ratio is shown for the duration when the giver was in contact with the  baton (i.e., from the beginning of the handover until  ). Similarly, the  for the duration when the receiver was in contact with the baton (i.e., from end of the handover). From the  /  /  is shown until the  plot, we can see that in the beginning of the handover,  when the giver was holding onto the baton alone, the force ratio was maintained fairly constant. During object transfer, the giver continued to maintain a roughly constant ratio, but toward the end, there was a great increase. The  /  /  ratio of the receiver, on the  other hand, started from a slightly lower value when the receiver first made contact with the baton. Toward the end of object transfer, the  /  ratio eventually rose to a stable value,  and remained roughly constant until the end of the handover trial. From these results, we can see that the giver’s and the receiver’s grip-force-to-load-force-ratios did not remain constant for the entire object transfer period. Therefore, the hypotheses stating that the giver’s and receiver’s grip-force-to-load-force-ratios would be approximately constant must be rejected.  29  3.4.2. Grip Force versus Load Force To examine how the giver and the receiver coordinate their grip forces with the load forces during object transfer, the experimenter plotted the grip force versus the load force of the giver (Figure 7A) and the receiver (Figure 7B). Note that in a handover, the baton starts in the giver’s hand, meaning that  starts from a high value, and  starts from zero.  Therefore, time travels right to left in Figure 7A, and left to right in Figure 7B. From Figure 7A and Figure 7B, we can see that both the giver’s and receiver’s grip force to load force relations appear to be fairly linear, and the curves from different conditions overlap each other, having approximately the same slopes. Linear grip-load relationships have also been observed by other researchers in pick-and-place tasks [17], [49], [57]. According to [17], a linear relationship in the grip-force-to-load-force plots implies constant ratios between the grip force rates and load force rates ( ̇ / ̇  and ̇ / ̇ ), and these ratios are equal to the  slopes of the lines. The curves for the giver in Figure 7A also appear to have positive intercepts on the vertical axis, but the curves for the receiver do not exhibit this behaviour.  Figure 7. Grip force versus load force plots of one pair of giver (A) and receiver (B). Both plots appear fairly linear. A positive intercept on the grip force axis is seen on the giver plot but not on the receiver plot. (Figure published in Chan et al. 2012 [50].) 30  Figure 8 plots the grip-load curves from all conditions for three giver-receiver pairs. Comparing the grip-load plots from different individuals, we can see that although the slopes of the curves remain fairly constant across different conditions for each person, they vary across participants. To characterize the relationship between the grip forces and load forces of the giver and receiver during object transfer, the experimenter performed linear regression on the data. Comparison of the regression variables across different conditions within each participant showed that both the slope and intercept did not vary significantly for the majority of participants. Comparison across individuals also showed that in general, there are no significant differences among the intercepts of the giver’s curves, and there are no significant differences among the intercepts of the receiver’s curves. The vertical intercept observed in the giver’s curve corresponds to the amount of grip force safety margin applied by the giver at zero load force and was found to be 2.3 ± 1.7 N.  Figure 8. Grip force versus load force plots for three different pairs of giver (A) and receiver (B) showing variations in the slope of the curves. (Figure published in Chan et al. 2012 [50].) 31  3.4.3. Transfer Time and Maximum Load Transfer Rate Figure 9 shows the average receiver load force, force,  , and receiver grip force,  , load transfer rate, ̇ , giver grip  , measured in each condition for a giver-receiver pair.  Table 2 shows the average transfer time, and time to maximum load transfer rate,  , maximum load transfer rate,  ( ̇ ),  , from all participants measured in each  condition; Table 3 shows the analysis of variance (ANOVA) results comparing these measures. Data analysis reveals that there were significant differences in  and  Figure 9. The average receiver load force, , load transfer rate, ̇ , giver grip force, , and receiver grip force, , in Light, Medium, and Heavy conditions. Dash lines show transfer period from one trial. 32  ( ̇ ) across different conditions. Post hoc analysis shows that  was  significantly longer in the Medium (p = 0.029) and Heavy (p = 0.002) conditions when compared to the Light condition. However,  in the Medium and Heavy conditions  did not differ significantly (p = 0.525). This suggests that although object weight, the relation is not linear as Significant differences were found for  increased with  appears to level off with heavier objects. ( ̇ ) across all conditions (p < 0.0005) with  ( ̇ ) increasing as object weight increased. This suggests that as the amount of  average  load to be transferred (object weight) increased, the receiver also tried to transfer the load more quickly. The time to maximum load transfer rate,  , however, did not vary  significantly across different conditions. Table 2. Measured transfer time, , maximum load transfer rate, ( ̇ time to maximum load transfer rate, , for each condition. ( ̇ Light Medium Heavy  470 ± 159 ms 500 ± 182 ms 523 ± 201 ms  ), and  )  19.00 ± 4.43 N/s 21.84 ± 4.80 N/s 25.14 ± 5.47 N/s  199 ± 123 ms 224 ± 150 ms 232 ± 167 ms  Table 3. Analysis of variance (ANOVA) results for comparing transfer time, , maximum load transfer rate, ( ̇ ), and time to maximum load transfer rate, . Significant results are marked with **. ( ̇ F p  F(1.70, 304) = 6.55 p = 0.003**  )  F(2, 358) = 104 p < 0.0005**  F(1.81, 324) = 2.64 p = 0.079  33  3.4.4. Maximum Excess Receiver Load (MERL) The maximum excess receiver load (MERL) measures the amount of excess load force experienced by the receiver toward the end of object transfer. Comparison of MERL% across conditions reveals no significant differences (F(2, 358) = 0.747, p = 0.474). The average MERL% was 2.36 ± 4.16% and a one-sample t-test show that this value is significantly different from zero (t(539) = 13.2, p < 0.0005). Figure 10 shows the histogram of MERL%. The distribution is centered around the average, and we can see that in most of the handovers (429 out of 540 handovers), MERL% was greater than zero. The positive MERL% measured suggests that most of the time (approximately 80%) the giver did not fully release the object until after sensing a small upwards pulling force from the receiver.  Figure 10. Frequency distributions of MERL% in all conditions. The first, second, and third quartiles were 0.5%, 2.5%, and 4.8%, respectively. This shows that most of the time MERL% was greater than zero, which means that the maximum load force experienced by the receiver during object transfer was greater than the object weight. (Figure published in Chan et al. 2012 [50].) 34  3.5. Discussion 3.5.1. Human Grip Force Control Strategy for Handovers Data from the experiment revealed that during object transfer, the giver and receiver do not simply maintain a constant grip-force-to-load-force-ratio throughout the entire process (Figure 6). Although the grip-load ratios were fairly constant for most of object transfer, they vary near the beginning (for the receiver) and toward the end (for the giver). Humans appear to regulate their grip force according to a different strategy than the one used for holding an object. Near zero load force, the giver applied an increased grip-load ratio, grasping the object much more tightly than necessary to prevent slip, whereas the receiver employed a much lower grip-load ratio, taking on the load of the object much quicker than building up the grip force. In handovers, humans do not simply regulate their grip forces proportional to the load forces. Upon further examining the grip versus load plots (Figure 7) it was found that the slopes of the giver’s and receiver’s curves were maintained fairly constant. This result is similar to what has been found in pick-and-place tasks [17], [57] – when a person picks up or places an object on a table, the resulting grip-load curves have approximately constant slopes. This indicates that humans may be employing similar strategies for both pick-and-place tasks and handovers. Section 3.4.2 reported that although the slopes of the grip-load curves are generally independent of object weight within each individual, there can be huge variations across individuals. The slope variations across different people may be attributed to different surface coefficients of friction between each person’s fingers and the baton. A lower coefficient of friction would require more grip force to prevent slip given the same amount of load. This explanation is consistent with the literature [17], [49], [57]. Despite the variations in the slopes, the positive intercept values of the giver’s curves did not differ significantly across the majority of participants. Most people appear to apply approximately the same amount for grip force safety margin at the point when  becomes zero.  35  To further explore the similarities between handovers and pick-and-place tasks, the experimenter compared the characteristics of the grip force and load force data from the handover experiment with literature data on pick-and-place tasks [16], [17] in greater detail to determine whether a) giving/releasing an object to another person is the same as placing/releasing an object on a table, and b) taking/receiving an object from another person is the same as taking/picking-up an object from a table. Comparison of the data reveals that handing over an object is different from placing an object, and receiving an object is different from picking-up an object. On the contrary, features observed in a giving task are more similar to a picking-up task, while features observed in a receiving task are more similar to a placing task. According to the literature, when picking-up an object, people enter a “pre-loading” phase in which they increase the grip force before lifting the object and taking on any load [16], [17]. This causes a high grip/load ratio near zero load force and appears as a nearvertical segment in a grip-load plot near the origin. These features do not appear in the receiver data, but are observed in the giver data. It appears that there is a “post-unloading” phase during which the giver maintains a positive amount of grip force even when the entire load of the object has been transferred to the receiver. The grip/load ratios in a receiving task and a placing task, on the other hand, do not have increased values and both ratios start at lower values near zero load force. Common features observed between giving tasks and picking-up tasks, and between receiving tasks and placing tasks may be a reflection of the similarities in relative tasks difficulties. Picking-up an object may be a more difficult task compared to placing an object since the person has to deal with more uncertainties. These uncertainties, related to object properties, include surface friction and object weight. If a person attempts to pick up an object with insufficient amount of grip force, the object will slip and drop. The person must rely on past experience and/or memory to predict and estimate the object’s properties for proper grip force control [17]. Therefore, when a person picks up an object, a “pre-loading” period is needed to build up the grip force safety margin. Placing an object, however, only requires the person to completely reduce the grip force once the object has made stable contact with the table. There is no risk of the object dropping past the table. 36  Similarly, handing over an object is a more difficult task compared to receiving an object. During a handover, the control of the speed of object transfer lies mainly with the receiver. The receiver controls how fast the object is being taken, while the giver must respond to the load force changes. The giver can only infer whether the receiver has control of the object through the absence of slip. If the giver misjudges and releases too soon, the object may be dropped in the process of the handover. Therefore, to ensure a safe object transfer, the giver appears to employ a “post-unloading” phase for increasing the safety margin near the end of object transfer. The receiver, on the contrary, simply needs to increase the grip force to take the object once it is in-hand and does not need to apply an increased grip/load ratio near zero load force.  3.5.2. Transfer Time and Maximum Load Transfer Rate Object weight affects handover difficulty. A heavier object increases the difficulty of the handover by requiring more effort to carry the object and more attention from the giver and receiver, since dropping the object leads to more severe consequences (e.g., damage and injury). Results show that when object weight increased from Light to Medium or Heavy, increased significantly and object transfer took longer. This increase in transfer time may be due to increased task difficulty. However,  appeared to level off and  did not increase significantly when the object weight increased from Medium to Heavy. This may indicate that handover task difficulties at these two weights are similar. However, as the weight of the object continues to increase, it can be expected that the task difficulty will once again increase. With heavier object weight, the  ( ̇ ) also increased significantly,  indicating that the receiver adjusts the rate of load transfer accordingly in order to maintain efficient object transfer times.  37  3.5.3. Maximum Excess Receiver Load (MERL) The average MERL% in human-human handovers was found to be 2.36%. Although this number is small, analysis confirmed that it is significantly greater than zero. This is a result of the giver delaying the release and continuing to hold on to the object even after all of the load has been transferred to the receiver. The experimenter interpreted the delay of release as a precautionary measure of the giver for ensuring a safe object transfer. A small amount of “pull” from the receiver is required to signal the release of the object and this appears to be a mutually established component of the haptic communication in handovers as it is consistently observed across different people.  3.5.4. Implications on Robot Handover Controllers The results from the human-human handover experiment reported herein provide a better understanding of the haptic interaction in human-human handovers. Based on these findings, this thesis proposes the following guidelines for designing safe, efficient, and intuitive robot handover controllers for handing over everyday objects. Object Transfer Period. Section 3.4.3 reported that the object transfer time, , was approximately 500 ms in all conditions, which agrees with the literature [44]. Based on this, it is recommended that the duration of human-robot object transfer should be approximately 500 ms. Object Safety and Load Transfer Speed. An examination of the giver’s grip-forceto-load-force-ratio in Section 3.4.1 showed that near the end of object transfer, the giver applies an increased grip/load ratio. Section 3.5.1 explained that this increase in the grip/load ratio corresponds to a “post-unloading” employed by the giver for ensuring a safe completion of the object transfer. In human-human handovers, there seems to be an implicit understanding that the giver should take on the responsibility of ensuring object safety. Therefore, as giver, a robot should ensure object safety. Examining the rate of load transfer, Section 3.4.3 reported that  ( ̇ ) increases with object weight. Section 3.5.1 pointed out  38  that the speed of object transfer is controlled mainly by the receiver. Given a greater amount of load to be transferred, a human receiver increases long  ( ̇ ) to prevent an inordinately  . Therefore, as receiver, the robot should be responsible for maintaining load  transfer speed. Grip Force Safety Margin at Zero Load Force. Section 3.4.2 showed that the giver’s grip-load curve exhibits a positive intercept, meaning that at zero load force, the giver still maintains a positive amount of grip force. This residual amount of grip force is the safety margin applied by the human giver for ensuring a safe completion of object transfer. According to this, a robot giver is recommended to apply a non-zero grip force even when the load force has reached zero. Complete Release of Object. Observing the maximum amount of load force the receiver experiences, Section 3.4.4 showed that the average MERL% is typically around 2.5%. Section 3.5.3 pointed out that this is another precautionary measure employed by the giver for ensuring a safe object transfer. The giver does not release the object until a slight “pull” is sensed. Therefore, a robot giver should relinquish the object only after a negative load force is sensed.  3.6. Summary Through the study presented in this chapter, this thesis has provided a better understanding of the haptic interaction in human-human handovers. During a handover, the giver and receiver employ a similar grip force control strategy. However, the role of ensuring object safety and maintaining load transfer speed are assigned separately to the giver and the receiver, respectively. It appears that there is an established mutual understanding of this division of responsibilities among people, and observations across subjects are consistent. Comparison of our findings to the literature revealed that although there are similarities between control strategies use for pick-and-place tasks and for handovers, there are significant differences. In fact, grip force and load force features of a giving task are more similar to a placing task, whereas features of a receiving task are more similar to a placing 39  task. This indicates that using a pick-and-place controller is inadequate for handover tasks, and that a specific handover controller is required. Based on the results presented in this chapter, the following chapter presents the design of a novel robot handover controller for allowing robots to handover objects to people in a safe, efficient, and intuitive manner.  40  4. Robot Handover Controller Design Chapter 3 reported that during a human-human handover, the giver actively regulates the grip force such that an approximately linear relationship is observed between the grip force and load force. Furthermore, the giver takes on the responsibility of ensuring the object’s safety during the handover. Chapter 3 concluded with several guidelines for designing safe and natural human-robot handovers. Based on the results from Chapter 3, this chapter presents a novel handover controller design. The design follows the guidelines provided by Chapter 3 and allows a robot to safely, efficiently, and intuitively hand over objects to people by mimicking human grip force control. This chapter documents the design of the handover controller in Section 4.13. The novelty of the design lies within a load-force-dependent grip control function. The grip control function contains three parameters that allow the controller to be tuned for different applications. Section 4.2 discusses the selection of the parameter values and how each parameter influences the resulting handovers. Section 4.3 outlines some of the advantages of the proposed controller, and Section 4.4 provides a summary of the chapter.  4.1. Handover Controller Design This section reports on the grip force control function and describes the parameters that modulate the controller behaviour.  3  The controller design documented in this chapter is filed in a provisional patent application:  E. A. Croft, W. P. Chan, C. A. C. Parker, and H. F. M. Van der Loos, “Method and System for Releasing an Object Held by a Robotic End Effector,” U.S. Patent Application 61/696,671, 2012 (Provisional).  41  4.1.1. Grip Force Control Function Figure 11 shows the proposed grip force control function for a robot giver. The horizontal axis,  , is the measured load force at the end effector, and the vertical axis,  , is  the grip force to be applied. Leading up to the handover task, the robot is assumed to have the object stably grasped by its end effector. During this stable grasp, the robot experiences a load force of of  , corresponding to the weight of the object, and applies an initial grip force  . The initial grip force,  both the object load,  , must be sufficient to prevent slip, and is thus dependent on  , and the coefficient of friction between the object surface and the  grasping fingers. It is recommended that  include a factor of safety in addition to the  minimum amount required to prevent slip [17]. The proposed design assumes that  has  been appropriately determined by an available grasping algorithm. During object transfer, the robot regulates the amount of applied grip force, load force,  , according to the amount of measured  , following the linear relationship shown in Figure 11. As the load force reaches  zero, the robot maintains a small amount of positive grip force, force safety margin. Herein, the parameter  , which serves as a grip  will be referred to as the zero load grip force.  Once the robot senses a small amount of negative load force,  , the robot then releases  Figure 11. Proposed grip force control function. This function mimics human grip force control strategy. During object transfer, the applied grip force, , is regulated according to the experienced load force, , in a linear fashion such that the grip force is gradually reduced as the load force decreases. The object is completely released after a slight upwards pulling force of is sensed. 42  the object. This is the release force threshold and its magnitude corresponds to the amount of pulling force required from the receiver (This Chapter will discussion  further in  Section 4.1.3 and Section 4.2). Referring to Figure 11, the grip force control function can be expressed as: (4) where  is the slope of the function and can be calculated as: (5)  Combining Equations (4) and (5), yields: (6) During object transfer, the controller uses Equation (6) to compute the proper amount of grip force to be applied, based on the amount of load force.  43  4.1.2. Measured Load Force and Dynamic Effects Acceleration data from the human-human handover study presented in Chapter 3 showed that the object transfer phase is generally quasi-static – i.e., during object transfer, accelerations of the object are small enough such that dynamic forces are negligible. In quasi-static situations, all forces acting on the object sum up to approximately zero. Therefore, we can write: (7) where  and  are the vector forces (bold fonts denote vectors) applied by the  giver and receiver, respectively, and  is the mass of the object,  is gravitational acceleration,  is the overall acceleration of the object (Figure 12A). The load force experienced by  the giver,  , and the load force experienced by the receiver,  opposite direction of the applied forces and  and  , are equal to and in , respectively (i.e.,  ). Therefore Equation (7) can be re-written as: (8)  Rearranging Equation (8), we get an expression for  : (9)  Data from the human-human handover experiment showed that during object transfer, planar forces (forces in the x and y directions) applied to the object are generally much smaller than  Figure 12. Free body diagrams of the object in cases when A. acceleration is negligible, and B. acceleration is non-negligible. 44  the vertical forces (forces in the z direction), which means that  and  lie mostly in  the vertical direction (along the z axis). Therefore, Equation (9) can be simplified into a scalar equation in the z direction: (10) Equation (10) shows that in quasi-static situations,  reflects the amount of object  load supported by the receiver. Thus, a robot can infer when the receiver has gained control of the object through measurement of  , and this is the underlying principle of regulating  grip force according to the load force. However, in certain situations non-negligible dynamic forces may arise due to large accelerations (Figure 12B), e.g., a mobile robot travelling on uneven ground, or the robot arm experiencing a collision. In such cases, Equation (9) becomes: (11) and the vertical load force becomes: (12) where  is the acceleration of the object in the vertical direction. Equation (12) shows that  accelerations in the vertical direction,  , may cause  to decrease even when the receiver is  not supporting any part of the object’s weight. Therefore, large vertical accelerations could cause the robot to misjudge whether the receiver has taken control of the object or not. To prevent dynamic effects from triggering false handovers, the controller will ensure that the acceleration of the object (or end effector) is sufficiently small before releasing the object. Since planar accelerations (accelerations in the x and y directions) could also cause the object to slip out of the robot’s gripper without the robot releasing the object, the controller will monitor acceleration in three-dimension. If acceleration (in any direction) exceeds a certain limit, the controller will suspend the handover and take appropriate measures to ensure safety of the object. The next section elaborates on this topic.  45  4.1.3. Handover Controller Flow Chart Figure 13 shows a flow chart for the handover controller. Prior to object transfer, the grip force control function  is computed according to Equation (6). During each real-  time control cycle, the controller first checks to see if the acceleration,  , of the object is  within a tolerance value α. If the acceleration is greater than α, the situation will be deemed as unsafe for object transfer, and the controller will continue to hold on to the object. At this time the controller may simply apply a grip force of  , or execute an alternative grasp  algorithm to ensure a stable grasp around the object. On the other hand, if the acceleration is less than α, the controller will proceed with object transfer and adjust the robot’s grip force according to the grip control function, the release force threshold,  . While the measured load force remains above  , and the acceleration remains less than α, the controller  continues to regulate the grip force according to  . Once the load force drops below ,  the controller then opens the gripper to release the object, and the object transfer is completed.  Figure 13. Handover controller flow chart. 46  4.2. Controller Parameters As discussed in Section 4.1, the proposed controller design has three tuneable parameters, summarized in Table 4. The following subsections examine how each parameter can be expected to affect the resulting handovers as well as limiting factors on the selection of the parameter values. Table 4. Tuneable parameters of the proposed handover controller. Controller Parameter  Description Zero load grip force Release force threshold (less than zero)  α  Acceleration tolerance  4.2.1. Zero Load Grip Force The zero load grip force, load. With a higher  , is the residual amount of grip force applied at zero  , the controller applies a higher amount of grip force at zero load,  permitting the robot to more securely hold on to the object near the end of object transfer. However, a high  would also mean that when  reaches , the robot will have to reduce  its grip force to zero from a higher value. If the robot cannot reduce its grip force fast enough due to physical limitations, object release may be delayed past the intended release force threshold, and the receiver may experience a “pull” greater than  (this effect is confirmed  in Section 5.5.4). Theoretically, at zero load force, the amount of grip force required to prevent object slip is zero; any amount of grip force applied at zero load is purely precautionary. However, Section 3.4.2 showed that human givers typically apply a small amount of grip force (~2 N) at zero load, and this small amount of grip force serves as a safety margin for ensuring a safe completion of object transfer. Therefore, the value of selected for the controller should be small but non-zero. Given a fixed object weight, load grip force,  , and a corresponding initial grip force,  , the zero  , determines the slope of the grip force control function according to  Equation (5). The lower the value of  , the greater the slope of  , and the faster the 47  robot will have to be able to vary its grip force. Therefore,  is indirectly limited by the  response time of the robot’s grip force controller. Section 3.4.3 reported that the object transfer phase of a natural human-human handover is approximately 500 ms; the chosen value of  should allow the robot to smoothly traverse  within this timeframe.  4.2.2. Release Force Threshold The magnitude of the release force threshold, i.e.,  , corresponds to how much the  receiver needs to “pull” on the object before the robot releases. According to Section 3.5.4 the release force threshold should be negative in order to ensure a safe completion of object transfer. While a smaller magnitude of allows the receiver to take the object with less force, a greater magnitude makes the object less likely to be dropped due to accidental collisions. A greater release force threshold magnitude can also be used to compensate for inaccurate acceleration measurements. If a robot’s acceleration sensors are noisy, the release force threshold can be increased to provide better resistance to dynamic forces. Referring to Figure 11, note that once the zero load grip force,  , is chosen, the  grip force control function is defined according to Equation (6), and the horizontal intercept of the function is fixed. When the value of  decreases below the horizontal intercept,  becomes negative. A robot cannot apply a negative amount of grip force, and therefore, the value of  cannot be smaller than the horizontal intercept. This means that  imposes a restriction on the release force threshold value: (13) Equation (13) can be interpreted in two ways: 1) given a  value, the release force  threshold, , must not be smaller than the horizontal intercept, or 2) given a load grip force,  value, the zero  , must be high enough to satisfy Equation (13).  The physical properties of the object being handed over also impose limits on the release force threshold. When handing over a delicate object, if the receiver is required to pull very hard to take the object, the object could be broken during object transfer. Therefore, 48  if an object is fragile or delicate, the magnitude of the release force threshold should be lower. Furthermore, a large magnitude of  would also result in high jerks at the point of release  (see video demonstration [34] of Bohren et al.’s implementation [32]). Therefore if it is undesirable to have the object shaken violently (e.g., when handing over a bottle of soda or a glass of water), the release force threshold magnitude should also be kept low.  4.2.3. Acceleration Tolerance The acceleration tolerance, α, determines how robust the controller is to dynamic forces that may arise due to robot movements or unexpected collisions. The smaller the value of α, the smaller the chance of dropping the object due to dynamic disturbances. Results from the human-human handover experiment showed that object transfer in the study was generally quasi-static; the average RMS z-acceleration measured during object transfer was less than 0.4 m/s2. Therefore, in similar settings, the value of α can be chosen to be as close to 0.4 m/s2 as possible to provide the most amount of resistance to dynamic disturbances. However, practical values of α will be limited by the accuracy of the sensors; the value of α must remain above noise levels. Furthermore, sensor lag time can also affect the performance of the controller. If the robot’s accelerometers have a greater lag time than the load force sensors, the measured  value may change before the acceleration is detected. Therefore,  when choosing the value of α, sensor lag time should be taken into consideration.  4.3. Discussion It is important to note that the grip force control function presented in Section 4.1.1 is independent of time. Furthermore, during a handover, the robot is allowed to traverse the grip force control function in both directions. These features make the controller robust to small load force perturbations and provide several advantages: Time Independence. Section 3.4.3 reported that in human-human handovers, transfer time and maximum load transfer rate vary with object weight; this implies that the timing of 49  object transfer varies with different objects. The time-independent feature of the proposed grip control function allows the robot to automatically adapt to timing variations among handovers with different objects. Furthermore, Section 3.5.4 stated that the responsibility of controlling the rate of load transfer lies with the receiver; the proposed grip control function also allows the receiver to decide when to begin object transfer and to control the overall timing of the handover. Object Weight Changes. The grip force control function, load forces greater than the initial weight of the object (i.e.,  , is valid even for ). Thus, in cases where  the weight of the object increases (for example, water being poured into a held cup), the controller will be able to increase the grip force accordingly to maintain a stable grasp, provided that the slope of  is not too small. With other controllers where grip force is  held constant, if the weight of the object increases, the load force will eventually exceed the slip threshold, and the object will drop. However, the proposed controller design actively monitors the load force and regulates the applied grip force appropriately to prevent the object from slipping. Receiver Intent Changes. Since the controller design permits traversal of the function in both directions, even if the receiver chooses to abort the handover partway into object transfer, the robot will be able to recover. After the receiver has initiated object transfer and has begun taking on part of the object load, if he/she then releases the object suddenly, due to error or change of intention, the proposed design will allow the robot to regain a stable grasp by increasing the grip force as the load force on the robot increases. Of course, once the release force threshold, , has been reached, the controller will release the object and responsibility for the object shifts entirely to the receiver.  4.4. Summary This chapter proposed a novel handover controller design that allows a robot giver to mimic human grip force control when participating in robot-to-human handovers. The design features a load-force-dependent grip force control function and enables the robot to  50  appropriately regulate grip force based on measured load force during handovers. As discussed in Section 4.3, the proposed design can adapt to variations in handover timing and is robust to several types of unexpected changes in handover dynamics. Section 3.5.4 stated that in order to achieve safe and natural handovers, the giver should provide a grip force safety margin, such that a positive amount of grip force is applied at zero load force, and should release the object after a negative load force is experienced. The proposed controller complies with these guidelines by providing a positive zero load grip force and a negative release force threshold. By tuning the zero load grip force and release force threshold values, the controller can also be adjusted to suit different applications. To verify the proposed handover controller design, the next chapter documents an implementation of the controller, as well as a user study conducted to compare different tunings of the parameters.  51  5. Robot-Human Object Handover Study The previous chapter documented a novel design of a robot handover controller. To verify the design, this chapter provides an implementation of the controller on a Willow Garage PR2 robot. A user study comparing four different tunings of the controller shows that the proposed design produces safe, efficient, and intuitive handovers. Results show that the proposed load-force-dependent grip control function yields measureable improvements over existing constant-grip controllers, and among the four controllers tested, participants preferred the controller that produces the most human-like handovers. This chapter is organized as follows: Section 5.1 documents the implementation of the proposed controller on a Willow Garage PR2 robot. Section 5.2 discusses four different tunings of the controller. The research questions and hypotheses are stated in Section 5.3, and the experimental method is explained in Section 5.4. Section 5.5 presents the experimental results, followed by discussion in Section 5.6. Section 5.7 summarizes the work presented in this chapter.  52  5.1. Handover Controller Implementation 5.1.1. Hardware Platform The experimenter implemented the handover controller described in Chapter 4 on a Willow Garage PR2 robot (Menlo Park, CA, USA). Figure 14 shows the PR2 used for this implementation. The PR2 has an anthropomorphic upper body mounted on an omnidirectional moveable base. It has two seven-degree-of-freedom (DoF) arms, and each arm is equipped with a two-finger, one-DoF parallel-jaw gripper. Each finger is covered with a protective layer of silicone rubber, which provides a compliant, non-slip surface for grasping.  Figure 14. Willow Garage PR2 robot used in the robot-human handover experiment. 53  5.1.2. Load Force Sensing The PR2’s controller provides torque measurements for each of its joints at a rate of 1 kHz. The robot’s forearm and gripper are modelled as a two-link mechanism for calculating the load force (Figure 15). In this implementation, the robot’s forearm and gripper are positioned such that both the wrist joint’s axis of rotation and the gripper’s axis of actuation lie in the plane parallel to the ground. In static conditions, the magnitude of the torque applied by the motor at the wrist joint, , is equal to the sum of the torques due to the weight of the gripper,  , and the load force at the gripper,  : (14)  where  is the length from the wrist joint to the gripper’s center of mass, and  is the  length from the wrist joint to the gripper tool center. Rearranging Equation (14) yields an expression for  in terms of : (15)  Figure 15. Two-link model of robot forearm and gripper for calculating load force. 54  Initial tests showed that the joint torque measurements at the PR2’s wrist, , exhibited hysteresis and drift. Typical torque measurements varied up to 30% across multiple readings. To account for errors in the torque measurements, Equation (14) was modified to include an error term  : (16)  Thus, the expression for  becomes: (17)  where (18) is the torque measurement offset. To calculate finding the value of  from Equation (17), we need to find . The first method is to determine  the handover. At that time, be the offset,  . There are two methods for before picking-up an object for  is equal to zero, and therefore, the measured torque value will  . The second method is to measure the torque value after the object has been  picked up. Given the weight of the object,  can be calculated using Equation (17). This  implementation of the controller on PR2 uses the second method. Since  is measured at a  time closer to the handover in the second method, this approach gives a smaller error due to drift and hysteresis when calculating  . In the experiment,  was calculated one second  after the robot began moving the object towards the receiver, immediately prior to starting the handover controller.  5.1.3. Grip Force Control PR2’s fingers are equipped with capacitive pressure sensing pads operating at 24.4 Hz. Controlling grip force using these pressure sensors results in a response time of approximately 100 ms. In comparison, the duration of object transfer in a typical humanhuman handover is in the order of 500 ms. Initial implementation showed that closed-loop 55  control using the pressure sensors was not fast enough for the handover controller design. Therefore, to achieve a faster response time, an open-loop, position-based grip force control approach was used; since PR2’s position control operates at 1 kHz, a position-based approach for grip force control permits a much faster response time. The gripper of the robot is approximated as a spring with a certain stiffness. When the gripper squeezes onto an object, the applied grip force,  , can be modelled by the equation: (19)  where  is the gripper position (the distance between the fingers),  is the gripper  displacement measured from the zero force position where the fingers are just touching the object, and the function,  , is the combined stiffness of the gripper and the object.  The stiffness is expressed as a function of  since it may not necessarily be constant.  Equation (19) can be simplified into: (20) where  . Therefore, if we can find the function  to  model the applied grip force, we can control grip force by controlling the gripper position. To find  , the gripper was calibrated against an ATI (Apex, NC, USA)  Mini45 6-axis force/torque sensor. The ATI force/torque sensor was placed between PR2’s gripper fingers, and the resulting grip force was measured while PR2 varied its gripper position. A second-order regression modelling  yielded an r-squared value of  greater than 0.99; Table 5 shows the regression variables. Performing position-based grip force control using this quadratic model reduced the response time to approximately 30 ms. Table 5. Coefficients for the robot grip force model . Coefficient  Value 14.67 x 106 N/m2 15.68 x 103 N/m  56  Note that the measured  is the combined stiffness of the gripper and the  object. The stiffness values of the ATI and the baton used for the study are assumed to be much greater than the stiffness of the gripper itself. Therefore,  can be used as a  good approximation of the gripper’s stiffness. Results from the experiment show that using this position-based grip force control method gives a precision of ~5 N. Since the minimum difference in zero load grip force,  , between any two controllers tested in the experiment  was 10 N (see Section 5.2), a precision of ~5 N was sufficient for this study.  5.1.4. Acceleration Measurements PR2’s grippers are equipped with accelerometers, and this implementation used these sensors to ensure that the object’s acceleration is within the acceleration tolerance, α, at the time of object release. The accelerometers measure accelerations up to 78 m/s2 and have a nominal resolution of 0.15 m/s2 [58]. Acceleration data is sampled at 3 kHz, with successive measurements provided to the controller in groups of three at a rate of 1 kHz. PR2’s controller provides a filtered reading of the grippers’ accelerations using a first-order Chebyshev band-pass filter with cut-off frequencies at 5 Hz and 1000 Hz.  57  5.2. Controller Tunings The experimental study compared four different tunings of the proposed controller. As a first step toward characterizing the novel controller, this study addressed quasi-static object transfers, focussing on the two parameters of the grip control function, namely, the zero load grip force,  , and the release force threshold, . Figure 16 shows four different versions of  the grip force control function presented in Section 4.1.1, each with different values for and/or . In this study, the initial grip force  was set to be 30 N. The value for the initial  grip force was selected manually such that the object would not easily slip out of the robot’s gripper. The acceleration tolerance, α, was set at 1 m/s2 for all controllers. This value was chosen experimentally to be as small as possible while remaining above the PR2’s accelerometer noise level. The tunings of the four controllers were as follows: Controller A (Human-like controller,  ): This controller was  tuned to produce human-like handovers. The zero load grip force and release force threshold of this controller were initially set to the values found in human-human handovers ( ). The experimenter then manually increased  and the magnitude of  to  account for sensor measurement errors so that the robot would not accidentally drop the object when handing it over.  Figure 16. Four different tunings of the handover controller compared in the robothuman handover study. 58  Controller B (Balanced controller,  ): The tuning of  Controller B aims to balance handover smoothness and object safety. To make the controller more robust to collisions and arm movements, the experimenter increased the release force threshold from Controller A. Different release force threshold values were evaluated using a displacement test. In the displacement test, the experimenter displaced the gripper by pushing it down a vertical distance of  , while the gripper was holding onto an object, and then  released it; since the PR2’s arms are impedance-controlled, this induces an oscillating motion. The controller passed the test if the object did not drop before the arm stopped oscillating. Using Bohren et al.’s displacement-based handover controller [32] as a benchmark,  was  chosen to be 1 cm (Bohren et al.’s controller was design to withstand 1 cm of displacement). The release force threshold value,  , for Controller B was determined by gradually  decreasing until the controller passed 10 out of 10 trials of the displacement test. Note that the zero load grip force,  , of Controller B was also increased to satisfy the restriction  given by Equation (13) in the previous chapter. Controller C (Constant-grip-force controller,  ):  Controller C was tuned to represent existing constant-grip-force controllers such as those found in [7], [31], [32]. Modifying Controller B, the zero load grip force,  , was set to  equal the initial grip force to produce a constant-grip-force control function. Using Controller C, the robot held onto the object with a constant amount of grip force and released it once a pulling force equal to the magnitude of the release force threshold was sensed. Controller C allowed the experimenter to compare the proposed control design with existing constant-grip-force designs. Controller D (Quick release controller,  ): According to  Section 3.5.4, a giver should release the object after a slight “pull” is experienced; that is, the release force threshold, , of the controller should be negative. This study explored the effects of releasing the object before the giver experiences a pulling force. Controller D was modified from the human-like controller (Controller A) by changing the release force threshold to a non-negative value. The release force threshold value for Controller D was manually adjusted so that the “pull” was no longer felt by the receiver. Controller D allowed  59  this study to explore the importance of a negative release force and the giver’s role of ensuring object safety as discussed in Chapter 3.  5.3. Research Questions and Hypotheses The main objectives of this study are to validate the novel handover controller design and to characterize the controller. The goal is to produce safe, efficient, and intuitive handovers, and the experimenter seeks to answer the following questions through this study: 1) Does the controller produce safe handovers? – The controller should allow the object to be safely transferred from the giver to the receiver without being dropped during the process. Furthermore, the human receiver should not feel that the object is in danger of being dropped during the handover. 2) Does the controller allow smooth and efficient object transfers? – The controller should allow the receiver to easily take the object without pulling very hard on it. From the receiver’s perspective, an excessive amount of force should not be required to take the object. 3) Does the controller enable intuitive handovers? – The receiver should be able to take the object from the robot without being given explicit instructions on how to do so. To answer these questions, the experimenter conducted a user study to compare the four different variations of the controller presented in Section 4.1. Existing studies on the reaching motions for handovers have shown that people prefer to work with robots that exhibit human-like characteristics [19], [20]. Therefore, the experimenter hypothesized that users prefer the human-like handover controller (Controller A). Furthermore, the experimenter also wanted to explore the significance of having a load-dependent grip control function and a negative release force threshold. These two features have been consistently observed in human-human handovers, and both are believed to be crucial parts of handovers. Therefore, the experimenter formulated the following hypotheses (superscript R denotes robot-human handover study): : Users  prefer the human-like controller (Controller A) among all four controllers. 60  : Users  are able to differentiate a constant-grip-force controller (Controller C) from a  grip-force-varying controller (Controller B). : Users  perceive a controller with a non-negative release force threshold (Controller D)  as less safe, and prefer a controller with a negative release force threshold (e.g., Controller A).  5.4. Method Twenty-four healthy adults (fifteen males and nine females) participated in a handover study with the PR2 (refer to Appendix C1 for consent form). In the experiment, the PR2 handed over a baton to the participants using the four different controllers, while the grip forces and load forces of the giver (the robot) and the receiver (the participant) were measured. Participants tested the controllers in pairs (e.g., Controller A and Controller D, Controller B and Controller C) and rated them comparatively in a survey following the use of each pair of controllers.  5.4.1. Experimental Design This experiment compared four pairs of controllers: (Controller A vs. Controller B), (Controller A vs. Controller C), (Controller A vs. Controller D), and (Controller B vs. Controller C). The first three pairs tested the hypothesis that the human-like controller is preferred over all the other controllers. The third pair also tested the hypothesis that a negative release force threshold is preferred. The last pair tested the hypothesis that a constant-grip-force controller is noticeably different from a grip-force-varying controller. To account for order of presentation effects, each pair of controllers was presented to the participants in both orderings. With four pairs of controllers, this resulted in a total of eight pairwise comparisons for each participant. To account for carryover effects across each pairwise comparison, the order in which each pairwise comparison was presented to each participant was determined using a Balanced Latin Square design. 61  5.4.2. Apparatus and Data Collection This experiment used the same baton and data collection setup as the one used in the human-human handover study documented in Chapter 3 for measuring the grip forces and load forces of the receiver and the robot (except that no accelerometers were attached to the baton or the receiver). For details of the apparatus and the data collection setup, refer to Section 3.2.1 and Section 3.2.2. The weight of the baton in this study was 447 ± 1 g, which is approximately the same as the weight of the baton in the Light condition (483 ± 1 g) in the human-human handover experiment. The weight difference resulted from the removal of the accelerometer on the baton.  5.4.3. Survey Design After each pairwise comparison, participants completed a survey with three questions. Each question asked the participant to compare a different aspect of the resulting handovers on a seven-point Likert scale. The three questions were: 1) Rate the likelihood that the object could have been dropped during the handover (1 – Likely to be dropped, 7 – Not likely to be dropped). 2) Rate how easy it was to take the object in the handover (1 – Very hard to take, 7 – Very easy to take). 3) Rate how much you prefer each handover (1 – Not at all preferred, 7 – Very much preferred). The survey also included a free-response question asking the participant to provide any comments they had about their experience with the robot handovers (A copy of the survey is included in Appendix C2).  62  5.4.4. Volunteer Recruitment Volunteers were recruited through advertisements posted around the University of British Columbia campus and on the social network Facebook. Calls for volunteers were also sent out via email to graduate students and department mailing lists of engineering, health sciences, and computer science departments of the University of British Columbia.  5.4.5. Experimental Setup and Procedure During the experiment, the PR2 was situated on one side of a table, and the participant was seated on the other side of the table. Two rectangular markers the size of the baton’s base indicated the starting positions for the PR2 and the receiver (Figure 17). The table setup was designed to account for PR2’s arm length. Before each handover, PR2’s right gripper was positioned directly above the marker in front of it, and the gripper was fully opened. The baton was placed on the marker in front of PR2, in between PR2’s fingers, ready to be grasped. When the handover began, PR2 closed its gripper around the bottom handle of the baton, picked the baton up, and reached over to the center of the table. Once  Figure 17. Table setup for the robot-human handover experiment. Two rectangular markers indicate the starting positions for the robot (giver) and the participant (receiver). 63  PR2 reached over, the participant then took the baton by the top handle, using the right hand with a precision grip (same type of grip as the one used in the human-human handover experiment). After object transfer, both PR2 and the participant brought their arms back to the starting markers to complete the handover. Participants were asked to maintain the vertical orientation of the baton and keep it off the table during each handover. Pilot testing of the experimental procedure was performed with eight subjects (not included in the reported set). Since one of the questions this study aims to explore is whether the proposed controller enables intuitive handovers, in pilot testing, the experimenter only told the participants to take the baton from the robot, and did not provide any explicit instructions on how to do so. The pilot test also only allowed four practice handovers, one handover per controller. However, after piloting with eight subjects, results showed that, although participants were able to complete object transfers with Controllers A and D, most participants were unable to reach the release force thresholds of Controllers B and C. During handovers with Controllers B and C, most participants would begin to take the baton. However, once their load force reached approximately 40%-50% greater than the object weight (which translates to approximately 2 N of pulling force), they would not continue to pull harder, and would simply continue to hold onto the baton. From pilot testing, the experimenter surmised that once the pulling force magnitude exceeded a certain threshold, the handover became unintuitive to users. To address this issue, the experiment script included explicit instructions to participants to pull as hard as they liked on the object until the robot let go. Furthermore, to permit the subjects to familiarize themselves with the robot and the handover controllers prior to data collection, participants performed practice handovers with all controllers twice, in random order. After this familiarization, the participants began the pairwise comparisons. In each pairwise comparison, PR2 first handed over the baton to the participant three times using one controller, and then another three times using the second controller. After two sets of three handovers with different controllers, each participant completed the survey questions. All twenty-four participants following this procedure in the full study were able to complete object transfer in all handovers.  64  5.4.6. Data Analyses To test the hypotheses stated in Section 5.3, the experimenter analyzed the survey answers for each pair of controllers compared by the participants. Two suitable tests for analyzing Likert scale data are the Sign Test and the Wilcoxon Signed-Rank Test. The Sign Test determines whether the number of differences in each direction between two samples are equal, and does not consider the actual measured values from the samples. The Wilcoxon Signed-Rank Test, on the other hand, determines whether the median of two samples are equal. The objective of this study was to determine if one controller was preferred over another. Participants provided relative ratings for each pair of controllers as a method of comparison, and the absolute value of the rating for each individual controller does not necessarily provide a meaningful measure of the controller. In such cases, the Sign Test is the more appropriate test [59]. Therefore, the experimenter used the Sign Test (significance reported at α ≤ 0.05) to analyze the survey responses for this experiment4. From the baton force measurements, the experimenter extracted the transfer time, , the maximum load transfer rate, maximum load transfer rate,  ̇  , the time from receiver contact to  , and the maximum excess receiver load percentage,  MERL%. These are the same measures used to characterize human-human handovers (refer to Section 3.3 for the definition of each measure). The experimenter also examined the grip force vs. load force plots of the giver (robot) and the receiver (participant). To compare between controllers, the experimenter calculated paired sample t-tests (significance reported at α ≤ 0.05) for each pair of controllers: (Controller A vs. Controller B), (Controller A vs. Controller C), (Controller A vs. Controller D), and (Controller B vs. Controller C). To characterize each controller, the experimenter combined all data from each controller for analysis. The measured MERL% and the vertical intercepts on the robot’s grip force vs. load force plots served as ground truth values for comparing with the parameters of each controller; MERL corresponds to the release force threshold magnitude, , and the vertical intercept of the robot’s grip-load plot corresponds to the zero load grip force,  .  4  Results from the Wilcoxon Signed-Rank Test were identical to the Sign Test in terms of significance found. Results of the Wilcoxon Signed-Rank Test are included in Appendix D.  65  5.5. Results 5.5.1. Survey Responses Table 6 through Table 8 show the analysis results for the survey responses. The negative differences column indicates the number of trials in which the participant gave a lower rating to the second controller compared to the first controller, while the positive differences column indicates the opposite. The ties column indicates the number of trials in which the participant gave identical ratings for both controllers. The Sign Test determines whether the number of negative differences is different from the number of positive differences. Table 6. Analysis results for comparing likelihood of dropping the object. Significant results are marked with **.  Controller A vs. Controller B Controller A vs. Controller C Controller A vs. Controller D Controller B vs. Controller C  Negative Positive Ties Differences Differences 11 19 18 14 19 15 34 2 12 7 7 34  Sign Test Result Z = -1.278, p = 0.201 Z = -0.696, p = 0.486 Z = -5.167, p < 0.0005** p = 1.0a  a. Binomial distribution used.  Table 7. Analysis results for comparing ease of taking the object. Significant results are marked with **.  Controller A vs. Controller B Controller A vs. Controller C Controller A vs. Controller D Controller B vs. Controller C  Negative Positive Ties Differences Differences 42 1 5 44 1 3 5 30 13 25 14 9  Sign Test Result Z = -6.100, p < 0.0005** Z = -6.261, p < 0.0005** Z = 4.057, p < 0.0005** Z = -1.601, p = 0.109  Table 8. Analysis results for comparing user preference. Significant results are marked with **.  Controller A vs. Controller B Controller A vs. Controller C Controller A vs. Controller D Controller B vs. Controller C  Negative Positive Ties Differences Differences 41 3 4 41 2 5 29 11 8 25 11 12  Sign Test Result Z = -5.578, p < 0.0005** Z = -5.795, p < 0.0005** Z = -2.688, p = 0.007** Z = -2.167, p = 0.030** 66  Likelihood of dropping the object. Analysis showed that participants’ ratings for Controllers A, B, and C regarding likelihood of dropping the object did not differ significantly. However, significantly more participants responded that the object was more likely to be dropped during handovers with Controller D when compared to handovers with Controller A. Ease of taking the object. Significantly more participants responded that the object could be more easily taken during handovers with Controller A, when compared to handovers with Controller B and Controller C. Significantly more participants also responded that the object could be more easily taken with Controller D compared to Controller A. There was no significant difference between Controller B and Controller C. Preference in controllers. There were significant differences in preference for handovers with Controller A over handovers with Controller B, Controller C, and Controller D. There was also a significant difference in preference for handovers with Controller B over handovers with Controller C.  67  5.5.2. Grip Force and Load Force Data Figure 18 plots the typical receiver load force, force,  , and receiver grip force,  , load transfer rate, ̇ , giver grip  , for all controllers from one participant. The plots  show all three handovers from one of the pairwise comparisons for each controller. Figure 18 also shows data from human-human handovers for comparison (human-human handover data from Light condition in Figure 9). From Figure 18, we can see that data from Controllers A and D are qualitatively similar to data from human-human handovers. However, in the ̇  plot of Controller D, we can see  that the maximum load transfer rate is much higher than what was observed in human-human handovers. From the  plot of Controller D, we can also see a small “bump” near the end  of object transfer, indicating a small amount of extra grip force being applied. In Figure 18, it appears that at the end of the handover, the receiver applies a different amount of grip force in the robot-human handover and the human-human handover; this merely indicates that the coefficients of friction between the receiver’s fingers and the object were different in the two occasions. Examining the  plots for Controller B and Controller C in Figure 18, we can see that  the peak load forces experienced by the receiver are much higher than those present in Controller A or Controller D. This is expected as Controllers B and C have higher release threshold magnitudes. The  plots of Controllers B and C show that the receiver has to  apply much higher grip forces to prevent the object from slipping while trying to take the object. The resulting object transfer times are also much higher for Controllers B and C. Examining the ̇  plots, we can see that during the latter half of object transfer, receiver load  rates reach larger negative values. This is the result of the maximum receiver load being much higher than the object weight during object transfer. Note that the receiver’s grip force and load force plots for Controllers B and C also exhibit higher variances.  68  Figure 18. The receiver load force, , load transfer rate, ̇ , giver grip force, , and receiver grip force, , for all controllers from one participant. The vertical dashed lines mark the object transfer boundaries for one of the robot-human handovers. Human-human handover data are also plotted for comparison. Data are aligned at the midpoint of object transfer. 69  5.5.3. Grip Force versus Load Force Figure 19 shows the grip force versus load force plots for all controllers from one representative participant. The plots of each controller show all three handovers from one of the eight pairwise comparisons. The robot’s grip-load plots show a distinct type of handover for each controller. Both Controller A and Controller D produce handovers with small zero load grip force,  . However, Controller D has a near-zero release force threshold, ,  whereas Controller A has a slightly greater Controller B produces handovers with a higher  magnitude comparing to Controller D. magnitude and  , and Controller C  produces handovers with fairly constant grip forces. The average measured initial grip force of the robot was 17.9 ± 5.4 N for all controllers.  Figure 19. Grip force versus load force plots for all controllers from one participant during object transfer. During object transfer, the giver moves down and left along the curve, while the receiver moves up and right. 70  Examining the receiver’s grip-load plots, we can see the plots are close to linear with similar slopes for all controllers. The main difference among the plots for each controller is the maximum amount of load force experienced by the receiver (i.e., how far to the right the receiver curves extend), which depends on the different release force thresholds. Examining the receiver’s grip-load plot for Controller D more closely, we can also see that the slope of the curves appears to increase near the end of the transfer (right side of the plot). This corresponds to the small amount of extra grip force applied near the end of object transfer noted previously in Figure 18.  5.5.4. Zero Load Grip Force and Maximum Excess Receiver Load (MERL) A linear regression analysis on the robot’s grip force versus load force data determined the actual achieved zero load grip forces and release force thresholds. Table 9 shows the controller input and measured values for MERL%. Note that by definition, MERL% is equal to  and , as well as the measured divided by the object weight.  Table 9. Measured zero load grip force, release force threshold, , and maximum excess receiver load percentage, MERL%, for the four controllers. Measured Controller  Measured  Controller  Measured  Inputs  Values  Inputs  Values  Controller A  10 N  4.0 ± 2.3 N  -2 N  -0.47 ± 0.50 N  10.7 ± 11.4%  Controller B  20 N  8.8 ± 3.5 N  -5 N  -2.59 ± 1.01 N  59.0 ± 23.0%  Controller C  30 N  14.3 ± 6.0 N  -5 N  -3.52 ± 1.60 N  80.4 ± 36.5%  Controller D  10 N  1.9 ± 3.5 N  +2 N  -0.04 ± 0.14 N  0.8 ± 3.2%  Table 10 shows the analysis results for comparing  MERL%  and MERL%. The  measured for Controller A was significantly smaller than the values for Controller B and Controller C, and was significantly larger than the value for Controller D. The measured 71  MERL% for Controller A was also significantly smaller than the values for Controller B and Controller C, and was significantly larger than the value for Controller D. Furthermore, a t-test showed that the  measured for Controller B was significantly smaller than the value  for Controller C, and the MERL% for Controller B was also significantly smaller than the value for Controller C. These results confirm that, while the controller input and measured values for  and differed considerably due to load force sensing and grip force control  errors, all four controllers produced distinct types of handovers, as intended. Table 10. Analysis results for comparing zero load grip force, , and maximum excess receiver load percentage, MERL%. Significant results are marked with **. Controller A vs. Controller B t(141) = -17.1, p < 0.0005** Controller A vs. Controller C t(140) = -24.4, p < 0.0005** Controller A vs. Controller D t(143) = 6.8, p < 0.0005** Controller B vs. Controller C t(143) = -9.4, p < 0.0005**  MERL% t(141) = -27.9, p < 0.0005** t(140) = -23.9, p < 0.0005** (t(143) = 11.0, p < 0.0005** t(143) = -7.1, p < 0.0005**  5.5.5. Transfer Time and Maximum Load Transfer Rate The average transfer times,  , average maximum load transfer rates,  ( ̇ ), and average time to maximum load transfer rate,  , for each controller are  shown in Table 11. Table 12 shows the analysis results for comparing and  between the controllers. The average  shorter than the average  for Controller B and Controller C. Average  for  for Controller A, while  for Controllers B and C did not differ significantly.  The average Average  ( ̇ ),  for Controller A was significantly  Controller D was significantly shorter than the average average  ,  ( ̇ ) for Controllers A, B, and C had no significant differences.  ( ̇ ) for Controller D, however, was significantly higher than the average  ( ̇ ) for Controller A. Average  for Controller B and Controller C were  significantly longer than the average  for Controller A. However, there were no  72  significant differences between the average  for Controllers A and D, or between  Controllers B and C. Table 11. Measured transfer time, , maximum load transfer rate, and time to maximum load transfer rate, , for each controller.  Controller A Controller B Controller C Controller D  1082 ± 637 ms 1856 ± 1113 ms 1982 ± 1122 ms 620 ± 336 ms  ( ̇ ) 18.81 ± 7.98 N/s 19.08 ± 8.27 N/s 19.70 ± 8.36 N/s 30.38 ± 14.35 N/s  ( ̇  ),  389 ± 414 ms 633 ± 784 ms 651 ± 757 ms 408 ± 320 ms  Table 12. Analysis results for comparing transfer time, , maximum load ̇ transfer rate, ( ), and time to maximum load transfer rate, . Significant results are marked with **.  Controller A vs. Controller B Controller A vs. Controller C Controller A vs. Controller D Controller B vs. Controller C  t(141) = -8.8, p < 0.0005** t(140) = -12.3, p < 0.0005** t(143) = 8.5, p < 0.0005** t(143) = -0.2, p = 0.808  ( ̇ ) t(141) = -1.4, p = 0.178 t(140) = -1.9, p = 0.061 t(143) = -1.5, p = 0.131 t(143) = -8.4, p < 0.0005**  t(141) = -2.7, p = 0.008** t(140) = -3.8, p < 0.0005** t(143) = -1.9, p = 0.066 t(143) = -1.5, p = 0.131  5.6. Discussion 5.6.1. Survey Responses Analysis of the survey answers provided support for all three hypotheses stated in Section 5.3. Results from Section 5.5.1 showed that participants indicated significant preference for human-like handovers (Controller A) when compared to the other handovers tested (Controllers B, C, and D). Participants were able to differentiate between a constantgrip-force controller (Controller C) and a grip-force-varying controller (Controller B), and showed significant preference for the grip-force-varying controller. Furthermore, participants  73  also indicated that handovers with Controller D, which has a non-negative release force threshold, felt less safe. Comments from the participants provided more insight on each Controller. Participants frequently described handovers with Controller A using the terms “smooth” and “natural”. One participant commented that Controller A “let go of the object much more easily”, which made the handovers “feel smoother, more natural”. Participants also said that handovers with Controller A felt “easy and effortless”. Another participant expressed that handovers with Controller A were “way more natural” and that “[i]t felt like… taking an object from a human”. These comments show that Controller A produced not only force measurements similar to those observed in human-human handovers, but also smoother and more natural handovers according to the users’ perspectives. Comments regarding Controllers B and C generally indicated that the amount of force required to take the object was too high. Participants expressed that “[n]either (Controller B or C) [they] liked very much” since they had to “yank [the object]… more than was necessary”, and it felt like they “had to wrestle the robot for the [object]”. Furthermore, the high release force threshold magnitudes caused frustration. One participant commented: “[Controllers B and C] made me feel impatient as the robot was holding on too long after I grabbed the object.” Another participant also said: “[Controllers B and C] make me slightly frustrated when taking the object because of the robot’s tight grip and long delay before releasing the object.” The high release force thresholds also confused users about the robot’s intention to hand over the object, and this is especially the case with Controller C. One participant said: “[Handovers with Controller C] felt almost as if I were stealing the box from the robot and it didn’t want to give it to me.” Another participant commented: “It was extremely hard to take the object from the robot in [handovers with Controller C]. He (the robot) just wouldn't let go!” One participant also said that, for Controller C, it “felt like the robot had attachment issues with the block.” These responses show that holding onto the object too long may confuse the users about the robot’s handover intention and invoke negative feelings from users. Participants in general commented that handovers with Controller D felt less safe and that the object had a high chance of being dropped. Some participants commented that 74  “[Controller D] released the object too quickly” and that they “almost dropped the object” in some of the trials. Others also expressed that they felt “[t]he object was … handed over pretty carelessly” during handovers with Controller D. One participant explicitly mentioned that the receiver “must add a sudden and unexpected additional amount of force to the grip to prevent [the object] from being dropped.” This explains the increased amount of grip force observed in the  plot in Figure 18, and the increased slope of the receiver’s grip-load  curve towards the end of object transfer, as seen in Figure 19. Section 3.5.3 suggested that a small amount of receiver pull is used for signalling object release in a handover, implying that the controller release force threshold should be a negative value. Controller D, which has a non-negative release force threshold, was intended to explore the effects of omitting the receiver pull. As expected, participants felt that handovers with Controller D were less safe and at a higher risk of dropping the object. However, data show that in addition to having a MERL% close to zero, the  ( ̇ ) was  also substantially higher for Controller D. As a result, it is not clear which of the two factors caused the receivers to feel less safe about the handovers. Further investigation is required to confirm the importance of a negative release force threshold. It is worth noting that even though participants felt that handovers with Controller D were less safe and expressed worries of dropping the object, only one participant actually dropped the object in any of the handovers with Controller D. This observation shows that humans are able to react to unexpected situations within a very short amount of time and recover quickly.  5.6.2. Grip Force and Load Force Data The handover controller implementation on PR2 used an open-loop position-based approach for controlling the grip force in order to achieve a fast response time. Comparing to the commanded  of 30 N, Section 5.5.3 reported that the average initial grip force  achieved was 17.9 ± 5.4 N. Although the error in grip force control, which can be attributed to the stiffness difference between the baton handle and the ATI force/torque sensor used for  75  calibration (refer to Section 5.1.3), was as high as ~10 N, the precision was on the order of ±5 N, and this proved sufficient for implementing the controller design; Section 5.5.3 and Section 5.5.4 confirmed that indeed Controllers A, B, C, and D produce four distinct types of handovers as intended, and the robot’s grip-load plots in Figure 19 show that each controller produces repeatable handovers in terms of achieved zero load grip forces and release force thresholds. Figure 18 shows that the resulting grip force and load force plots from Controller A are highly similar to those of human-human handovers (Figure 9). This similarity confirms that Controller A produces human-like handovers. The measured  and MERL% values  reported in Section 5.5.4 were higher than the human-human handover values reported in Section 3.4.2 and Section 3.4.4. However, the  and MERL% values for Controller A are  within one standard deviation and two standard deviations of the human-human handover values, respectively. Comparing the plots for Controllers B and C to those of Controllers A and D in Figure 18, it is clear that Controllers B and C exhibit much longer transfer times and much higher peaks. The high  peaks indicate that the receiver needs to apply a large pulling force  to take the object. Furthermore, the plots for Controller B and Controller C also show much higher variance in the  and  data; this might indicate that the human receiver is not as  familiar or comfortable with the handovers produced by these two controllers. Figure 19 shows that the receiver’s grip-load curves in this study were very similar to those shown in Figure 7 for human-human handovers; this indicates that receivers use similar grip force control strategies for both human-human handovers and robot-human handovers.  5.6.3. Maximum Excess Receiver Load (MERL) The for both Controller B and Controller C were set to have a larger magnitude than those of Controllers A and D, and the measured MERL% for Controllers B and C were indeed higher that the MERL% for Controllers A and D. However, even though the  for  Controller B and Controller C were both set to the same value, Section 5.5.4 showed that the 76  actual measured MERL% were different for the two controllers; the MERL% for Controller B was approximately 20% less than the MERL% for Controller C. These results show that by lowering the  , we can reduce the MERL% while keeping the same. This difference in  the MERL% may be due to the speed limitations of the robot’s gripper, as discussed in Section 4.2.1. With a higher  , the robot needs to reduce its grip force to zero from a  much higher value at the moment when the release force threshold is reached; the response time of grip force control could cause a delay in object release. Thus, given the same  input  value, by gradually decreasing the grip force (Controller B), the proposed controller allows the receiver to more easily take the object compared to a constant-grip-force controller (Controller C). The pilot study discussed in Section 5.4.5 showed that without explicit instructions to pull on the object, most users will not pull harder than ~40%-50% of the object’s weight. The MERL% for Controllers B and C both exceed 50%, which explains why the release force thresholds were never reached in the pilot study. It appears that once the release force threshold magnitude exceeds ~40%-50% of the object’s weight the handover becomes nonintuitive to users. The high release force thresholds of Controller B and Controller C were tuned to provide better resistance against dynamic disturbances. Although the MERL% of Controllers B and C are well below the value of a displacement-based controller that provides roughly the same level of safety5, the release force thresholds of Controllers B and C might still have been higher than what would be necessary if the robot had better acceleration sensing capabilities. PR2’s acceleration measurements had an error of 0.15 m/s2, which is twice as large as the average RMS acceleration measured in human-human handovers. If more accurate acceleration measurements are available, the release force threshold can be lower without affecting the robustness to effects of dynamic forces. Section 5.2 explained that Controller D was designed to have a non-negative release force threshold to eliminate the “pull” experienced by the receiver at the end of object transfer. Section 5.5.4 showed that the achieved MERL% for Controller D was significantly  5  MERL% measured for the displacement-based controller used in [32] was as high as 500%.  77  less than that of Controller A and was close to zero; this further confirms that Controller D does produce handovers with quicker releases and eliminates the “pulling” force felt by the receiver.  5.6.4. Transfer Time and Maximum Load Transfer Rate While the average transfer times, longer than the average ̇  , for Controllers B and C were significantly  for Controller A, the average maximum load transfer rates, ̇  , did not differ significantly among Controllers A, B, and C. The average ̇  values for Controllers A, B, and C were similar to the  in the Light condition of  human-human handovers. Recall that the baton weight in this experiment was approximately the same as the baton weight in the Light condition in the human-human handover experiment, and that  ̇  increased significantly as the object weight increased in the ̇  human-human handover experiment. It appears that given the same object weight, remains the same, independent of the handover controller used. The differences in can be attributed to the increased  for Controllers B and C. With an increased release  force threshold, participants did not change the ̇  which  , but instead delayed the time at  was achieved, taking a longer time to build up the load transfer rate.  Results show that the ̇  ̇  ̇  for Controller D was significantly greater that the  for Controller A. Compared with human-human handovers, the  Controller D was ~50% higher. This increase in  ̇  ̇  from  is an effect of the non-negative  release force threshold. Controller D releases the object before the robot’s load force reaches zero, thus causing the object to “drop”, forcing a much higher ̇  also resulted in a reduced  ̇  . This increase in  . While the action of taking the object is mainly  controlled by the receiver, results from Controller D show that a giver can encourage a quicker object transfer by releasing the object slightly earlier. Even though Controller A, according to participants, produced the most human-like handovers among the four controllers, the average  was approximately twice as long 78  as the average  measured in human-human handovers. The increased  may  indicate that people exercise more caution when taking an object from a robot. Although object handover is a common task, most people are not familiar with interacting with a robot. Therefore, participants may be slowing down the handovers because they are uncertain about how the robot might behave. Results show that participants took longer to ramp up to the maximum load transfer rate, and the three times as long as the average  averages for all controllers appear to be two to in human-human handovers. Since the proposed  handover controller is time-independent, the controller allows the receiver to adjust the timing of object transfer to a comfortable speed. Once users become more familiar with handing over objects with the robot, users may reduce  and object transfer time will  automatically shorten without any modifications to the controller.  5.6.5. Recommendations The handover controller implementation on the PR2 demonstrates that the proposed design can be implemented on a robot with only load force sensing and grip force control capabilities. If acceleration measurements are also available, the controller can also filter out dynamic forces and make the handovers more robust to collisions. Furthermore, through comparing four different tunings of the controller, this study has allowed us to explore and verify our predictions as to how the zero load grip force and the release force threshold affect the resulting handovers. Based on the experimental findings, this section provides several recommendations regarding the controller parameters and hardware specifications of a robot implementing a handover controller. Grip Force Control. Initial use of PR2’s gripper pads for grip force control resulted in a response time of ~100 ms. Preliminary testing showed that this was too slow for implementing the handover grip control function. Upon switching to a position-based grip control method, the response time was reduced to ~30 ms, and this provided fast enough grip force control. Results in Section 5.5.3 showed that the resulting precision was ~5 N, and this was sufficient for achieving consistent handover results. Therefore, the experimenter  79  recommends that a robot’s grip force control should have a response time of less than 30 ms, but can have an uncertainty as high as 5 N. Load Force Sensing. The release force threshold (corresponding to the MERL) is a haptic cue for signalling object release during a handover. A robot must be able to recognize this cue. Section 3.4.4 reported that MERL% in natural human-human handovers is only 2.36%, which translates to ~0.1 N. However Section 5.5.4 showed that the measured release force threshold magnitude for the human-like controller (Controller A) was close to 0.5 N, and users still felt that the resulting handovers were smooth and human-like. Therefore, a robot’s load force sensors should have at least 0.5 N resolution. Release Force Threshold. The smaller the release force threshold magnitude, the easier it is for the receiver to take the object. However, the receiver pull at the end of object transfer is an integral part of the haptic communication in handovers. Therefore, while the release force threshold should be small, it should remain non-zero and negative. Based on the measured MERL% of Controller A (~10%) and the MERL% found in human-human handovers (~2%), the experimenter recommends the release force threshold to be between 2%-10% of the object weight. Furthermore, the pilot study showed that without explicit instructions, most people will not apply a pulling force greater than ~40%-50% of the object weight. Therefore, even if the release force threshold magnitude is increased to increase object safety or to compensate for noisy acceleration sensors, the release force threshold magnitude should not exceed 40% of the object weight in order to allow intuitive handovers. Zero Load Grip Force. The zero load grip force corresponds to the grip force safety margin applied by the robot. Section 3.4.2 reported that human givers on average apply a zero load grip force of 2.3 N and Section 5.5.4 reported that the measured zero load grip force for the human-like controller was 4.0 N. Based on these measurements, the experimenter recommends that the controller zero load grip force should be approximately 2 N to 4 N. The zero load grip force depends on how accurate the robot is able to measure the load force and control the grip force. If the robot has more accurate sensors, the grip force safety margin could be reduced and the zero load grip force could be smaller.  80  5.7. Summary This chapter has verified the novel handover controller design presented in Chapter 4 by providing an implementation on a PR2 robot. A user study comparing four variations of the controller showed that users prefer Controller A, which produces the most human-like handovers. Results have shown that the human-like controller produces safe handovers: thirty-two participants (including pilot subjects) performed a total of 576 handovers with Controller A and all 576 handovers resulted in safe completion of object transfer. Results have shown that the human-like controller produces efficient handovers: receivers can effortlessly take the object from the robot, and the amount of “pull” the receiver experiences is only a small percentage of the object weight. The pilot study results also showed that the human-like controller produces intuitive handovers: people were able to take the object from the robot without any explicit instructions. Since the proposed controller design functions with only load force sensing and grip force control functionalities, the controller can be readily implemented onto most existing robot platforms. Results presented in this chapter show that the proposed load-force-based controller provides a highly functional and practical approach for robot-human handovers.  81  6. Conclusion This thesis has presented a controller design that enables robots to safely, efficiently, and intuitively hand over objects to people. To allow robots to hand over objects to people naturally in unscripted situations, the approach adopted in this work was based on the premise that robots should emulate human behaviour while participating in handovers. In order to implement human characteristics onto robots, the first part of this thesis work involved conducting a study to investigate the dynamic interaction in human-to-human object handovers. The human grip force control strategy identified in this study served as a basis for the design of a novel handover controller. The latter part of this thesis provided an implementation of the proposed design, and a second study was conducted to test the controller. Experimental data verified that the proposed controller allows users to safely, efficiently, and intuitively take the object from the robot.  6.1. Human-Human Object Handover The purpose of the human-human object handover study, presented in Chapter 3, was to characterize the haptic interaction in handovers. In this study, the grip forces and load forces of the giver and the receiver during a handover were examined. Results showed that during object transfer, both the giver’s and the receiver’s grip forces were tightly coupled to their respective load forces, and an approximately linear relationship was observed. In addition, the giver assumed the role of ensuring object safety, while the receiver was in charge of maintaining the speed of load transfer. These findings provided a better understanding of the dynamic interaction in human-human handovers, and guided the general design of the handover controller presented in Chapter 4.  82  6.2. Robot-Human Object Handover The handover controller design, presented in Chapter 4, features a load-force-dependent grip force control function. Existing controllers apply a constant grip force during object transfer and determine object release based on displacement thresholds or force/torque thresholds at the gripper. The controller proposed in this thesis differs by allowing a robot to mimic human grip force control strategy during handovers through actively regulating the grip force in response to the measured load force. To verify the proposed design, the first part of Chapter 5 documented a specific implementation of the controller on a Willow Garage PR2 robot. This implementation demonstrated that the controller design is realizable on robots with only modest load force sensing and grip force control capabilities. The latter part of Chapter 5 presented a user study conducted to compare four different tunings of the controller. Results show that users prefer the human-like controller, which has both a lower zero load grip force and a lower release force threshold magnitude compared to the other controllers tested. Results also indicated that the release force threshold should be less than zero, meaning that the robot should only release after a slight “pull” from the receiver is detected. Otherwise, receivers may feel that the object is in danger of being dropped during the handover. In addition, users prefer the proposed controller design when compared to existing constant-grip-force controllers.  6.3. Contributions The contribution of this thesis is twofold: First, through the human-human object handover study, this thesis has identified the giver’s and receiver’s grip force control strategies, and discovered the specific roles of the giver and the receiver during a handover. These findings have provided a better understanding of the dynamic interaction in handovers, and have served as a basis for designing smoother and more natural robot-human handovers. Second, this thesis has presented a novel handover controller design that enables robots to mimic human behaviour when handing over objects to people. The design has modest hardware requirements and thus provides a practical solution for most existing robot platforms. A user study verified that the proposed controller allows a human receiver to 83  effortlessly, and intuitively, take objects from the robot, and that the design yields smooth and natural handovers, according to users’ ratings.  6.4. Future Work This work has focused on haptic communication as a first step toward understanding the dynamic interaction in handovers. However, often there are also other communication channels available to the giver and receiver during a handover, e.g., visual, gestural, or verbal channels. Humans utilize vision, gestures, and speech in other object manipulation tasks and cooperative scenarios such as pick-and-place and turn-taking [55], [60–62]. Therefore, it is likely that these channels are also being used for facilitating object transfer in certain situations. One possible extension of this thesis is to investigate the roles of these extra communication channels during handovers. Other factors in handovers to be considered include object property and context of the handover. Object properties, such as surface friction, size, shape, and attached value, may influence a person’s preference in the controller tunings. For instance, a valuable object may require a higher release force threshold to reflect additional care given. Similarly, people may also have different expectations for handovers in different contexts; existing studies have demonstrated the importance of considering etiquette in human-robot interaction [63– 65]. For example, in certain cultures, a giver is expected to use both hands when handing over an object to a receiver with higher social status [63]. Thus, different social contexts may require different handover approaches. In addition, as an initial approach to characterize the proposed controller, this work has considered quasi-static handovers; certain situations may involve handing over objects with high accelerations, e.g., when the receiver is in a hurry, he/she may not wait for the giver to complete the reaching motion before taking the object [28]. Further investigation is required to extend the application of the proposed controller to such high-acceleration handovers. The proposed handover controller design has focused on robot-to-human handovers. The inverse situation, where the robot is acting as the receiver, would be another important direction for future research. A controller for receiving objects from people should also 84  recognize the haptic communication in handovers; the knowledge provided by the humanhuman handover experiment presented in this thesis can be used for guiding the development of a robot controller for receiving objects from humans. In addition to enabling human-torobot handovers, a robot receiver controller, together with the robot giver controller, can also be used for robot-to-robot object handovers. A load-force-based handover approach would allow a robot to hand over objects to other robots without the need to establish complex protocols. This approach can be used for handing over objects between two individual robots, or between two arms of the same robot as well.  85  Bibliography [1]  P. Akella and M. Peshkin, “Cobots for the automobile assembly line,” in Proceedings of the International Conference on Robotics & Automation, 1999, pp. 728-733.  [2]  W. Wannasuphoprasit, P. Akella, M. Peshkin, and J. Colgate, “Cobots: A novel material handling technology,” in Proceedings of IMECE, 1998.  [3]  United Nations, “World Population Ageing: 1950-2050,” 2002.  [4]  J. A. Mcmullin, M. Cooke, and R. Downie, “Labour Force Ageing and Skill Shortages in Canada and Ontario,” 2004.  [5]  W. Fung, Y. Leung, and M. Chow, “Development of a hospital service robot for transporting task,” in Proceedings of the International Conference on Robotics, Intelligent Systems and Signal Processing, 2003, pp. 628-633.  [6]  Z. Yubo, W. Zixi, J. Linhong, and B. Sheng, “The Clinical Application of the Upper Extremity Compound Movements Rehabilitation Training Robot,” in Proceedings of the 9th International Conference on Rehabilitation Robotics, 2005, pp. 91-94.  [7]  Y. S. Choi, T. Chen, A. Jain, C. Anderson, J. D. Glass, and C. C. Kemp, “Hand It Over or Set It Down: A User Study of Object Delivery with an Assistive Mobile Manipulator,” in Proceedings of the International Symposium on Robot and Human Interactive Communication, 2009, pp. 736-743.  [8]  R. Wilcox, S. Nikolaidis, and J. Shah, “Optimization of temporal dynamics for adaptive human-robot interaction in assembly manufacturing,” in Proceedings of Robotics: Science and Systems Conference, 2012.  [9]  K. Yamazaki, A. Yamazaki, M. Okada, and Y. Kuno, “Revealing gauguin: engaging visitors in robot guide’s explanation in an art museum,” in Proceedings of CHI’09, 2009, pp. 1437-1446.  [10] Y. Iwamura, M. Shiomi, T. Kanda, H. Ishiguro, and N. Hagita, “Do Elderly People Prefer a Conversational Humanoid as a Shopping Assistant Partner in Supermarkets?,” in Proceedings of the 6th International Conference on Human-Robot Interaction, 2011, pp. 449-456. [11] M. Diftler, J. Mehling, and M. Abdallah, “Robonaut 2 - the first humanoid robot in space,” in Proceedings of the International Conference on Robotics and Automation, 2011, pp. 2178-2183. 86  [12] Natural Sciences and Engineering Research Council of Canada, “NSERC Investments in Manufacturing (2010-11),” 2012. [13] Robotics Trends’ News Sources, “Robotic Assistants May Adapt to Humans in the Factory,” Robotics Trends, 2012. [Online]. Available: http://www.roboticstrends.com/industry_manufacturing/article/robotic_assistants_may _adapt_to_humans_in_the_factory/?utm_source=RoboTrendsWeekly&utm_medium= email. [Accessed: 12-Sep-2012]. [14] D. Kara, “Thinking Twice On Rethink Robotics,” Robotics Business Review, 2012. [Online]. Available: http://www.roboticsbusinessreview.com/article/thinking_twice_on_rethink_robotics. [Accessed: 12-Sep-2012]. [15] MarketsandMarkets, “Service Robotics Market (Personal & Professional) – Global Forecast & Assessment by Applications & Geography (2012 – 2017),” 2012. [16] R. S. Johansson and J. R. Flanagan, “Coding and use of tactile signals from the fingertips in object manipulation tasks.,” Nature reviews. Neuroscience, vol. 10, no. 5, pp. 345-59, May 2009. [17] R. S. Johansson and G. Westling, “Roles of glabrous skin receptors and sensorimotor memory in automatic control of precision grip when lifting rougher or more slippery objects,” Experimental Brain Research, vol. 56, no. 3, pp. 550-564, 1984. [18] J. Monzée, Y. Lamarre, and A. M. Smith, “The effects of digital anesthesia on force control using a precision grip.,” Journal of neurophysiology, vol. 89, no. 2, pp. 672-83, Feb. 2003. [19] S. Shibata, B. M. Sahbi, K. Tanaka, and A. Shimizu, “An Analysis of The Process of Handing Over An Object and Its Application to Robot Motions,” in Proceedings of the International Conference on Systems, Man, and Cybernetics, 1997, pp. 64-69. [20] M. Huber, M. Rickert, A. Knoll, T. Brandt, and S. Glasauer, “Human-robot interaction in handing-over tasks,” in Proceedings of the International Symposium on Robot and Human Interactive Communication, 2008, pp. 107-112. [21] P. Basili, M. Huber, T. Brandt, S. Hirche, and S. Glasauer, “Investigating HumanHuman Approach and Hand-Over,” Human Centered Robot Systems, vol. 6, pp. 151160, 2009. [22] K. Koay, E. Sisbot, and D. Syrdal, “Exploratory studies of a robot approaching a person in the context of handing over an object,” in Proceedings of AAAI Spring Symposium: Multidisciplinary Collaboration for Socially Assistive Robotics, 2007, pp. 18-24.  87  [23] M. Cakmak, S. S. Srinivasa, M. K. Lee, J. Forlizzi, and S. Kiesler, “Human Preferences for Robot-Human Hand-over Configurations,” in Proceedings of the International Conference on Intelligent Robots and Systems, 2011, pp. 1986-1993. [24] M. Cakmak, S. S. Srinivasa, M. K. Lee, S. Kiesler, and J. Forlizzi, “Using Spatial and Temporal Contrast for Fluent Robot-Human Hand-overs,” in Proceedings of the International Conference on Human-Robot Interaction, 2011, pp. 489-496. [25] S. Kajikawa and T. Okino, “Motion planning for hand-over between human and robot,” in Proceedings of the International Conference on Intelligent Robots and Systems, 1995, pp. 193-199. [26] A. Agah and K. Tanie, “Human Interaction with a Service Robot: Mobile-Manipulator Handing Over an Object to a Human,” in Proceedings of the INternational Conference on Robotics and Automation, 1997, no. April, pp. 575-580. [27] E. A. Sisbot, L. F. Marin-Urias, X. Broquère, D. Sidobre, and R. Alami, “Synthesizing Robot Motions Adapted to Human Presence,” International Journal of Social Robotics, vol. 2, no. 3, pp. 329-343, Jun. 2010. [28] F. Dehais, E. A. Sisbot, R. Alami, and M. Causse, “Physiological and subjective evaluation of a human-robot object hand-over task.,” Applied ergonomics, vol. 42, no. 6, pp. 785-91, Nov. 2011. [29] A. Edsinger and C. C. Kemp, “Human-Robot Interaction for Cooperative Manipulation: Handing Objects to One Another,” in Proceedings of the International Symposium on Robot and Human Interactive Communication, 2007, pp. 1167-1172. [30] K. Nagata, Y. Oosaki, M. Kakikura, and H. Tsukune, “Delivery by hand between human and robot based on fingertip force-torque information,” in Proceedings of the International Conference on Intelligent Robots and Systems, 1998, pp. 750-757. [31] T. Deyle, M. S. Reynolds, and C. C. Kemp, “RFID-Guided Robots for Pervasive Automation,” IEEE Pervasive Computing, vol. 9, no. 2, pp. 37-45, Apr. 2010. [32] J. Bohren et al., “Towards Autonomous Robotic Butlers: Lessons Learned with the PR2,” in Proceedings of the International Conference on Robotics and Automation, 2011, pp. 5568 - 5575. [33] H. Nguyen, A. Jain, C. Anderson, and C. C. Kemp, “A clickable world: Behavior selection through pointing and context for mobile manipulation,” in Proceedings of the International Conference on Intelligent Robots and Systems, 2008, pp. 787-793. [34] Willow Garage, “Beer Me, Robot,” 2010. [Online]. Available: http://www.willowgarage.com/blog/2010/07/06/beer-me-robot. [Accessed: 13-Jul2012]. 88  [35] M. Simoneau et al., “Role of the feedforward command and reafferent information in the coordination of a passing prehension task.,” Experimental brain research, vol. 128, no. 1–2, pp. 236-42, Sep. 1999. [36] R. S. Johansson and G. Westling, “Signals in tactile afferents from the fingers eliciting adaptive motor responses during precision grip,” Experimental Brain Research, vol. 66, pp. 141-154, 1987. [37] B. Siciliano and O. Khatib, “Handbook of Robotics,” Springer, 2008. [38] J. Ulmen and M. Cutkosky, “A robust, low-cost and low-noise artificial skin for human-friendly robots,” in Proceedings of the International Conference on Robotics and Automation, 2010, pp. 4836-4841. [39] H. Yussof, J. Wada, and M. Ohka, “Object handling tasks based on active tactile and slippage sensations in a multi-fingered humanoid robot arm,” in Proceedings of the International Conference on Robotics and Automation, 2009, pp. 502-507. [40] J. R. Flanagan, J. Tresilian, and a M. Wing, “Coupling of grip force and load force during arm movements with grasped objects.,” Neuroscience letters, vol. 152, no. 1–2, pp. 53-6, Apr. 1993. [41] J. Hermsdörfer et al., “Moving weightless objects Grip force control during microgravity,” Experimental Brain Research, vol. 132, pp. 52-64, 2000. [42] A. M. Wing and J. R. Flanagan, “Anticipating dynamic loads in handling objects,” in Proceedings of the ASME Dynamic systems and Control Division, 1998, pp. 139-143. [43] R. S. Johansson and G. Westling, “Programmed and triggered actions to rapid load changes during precision grip,” Experimental Brain Research, vol. 71, pp. 72-86, 1988. [44] A. H. Mason and C. L. Mackenzie, “Grip forces when passing an object to a partner,” Experimental brain research, vol. 163, no. 2, pp. 173-87, May 2005. [45] Barrett Technology Inc., “BarrettWAM.” [Online]. Available: http://www.barrett.com/robot/products-arm.htm. [Accessed: 21-Sep-2012]. [46] Barrett Technology Inc., “BarrettHand.” [Online]. Available: http://www.barrett.com/robot/products-hand.htm. [Accessed: 21-Sep-2012]. [47] Robotiq, “Robotiq 3-Finger Adaptive Robot Gripper.” [Online]. Available: http://robotiq.com/en/products/adaptive-robot-gripper. [Accessed: 21-Sep-2012]. [48] Willow Garage, “PR2 Hardware Specs - Willow Garage.” [Online]. Available: http://www.willowgarage.com/pages/pr2/specs. [Accessed: 21-Sep-2012]. 89  [49] M. K. Burstedt, J. R. Flanagan, and R. S. Johansson, “Control of grasp stability in humans under different frictional conditions during multidigit manipulation.,” Journal of neurophysiology, vol. 82, no. 5, pp. 2393-2405, Nov. 1999. [50] W. P. Chan, C. A. C. Parker, H. F. M. Van der Loos, and E. A. Croft, “Grip Forces and Load Forces in Handovers: Implications for Designing Human-Robot Handover Controllers,” in Proceedings of the 7th ACM/IEEE International Conference on Human-Robot Interaction, 2012, pp. 9-16. [51] ATI Industrial Automation, “ATI Force - Torque Sensor- Mini45.” [Online]. Available: http://www.ati-ia.com/products/ft/ft_models.aspx?id=Mini45. [Accessed: 21-Sep-2012]. [52] R. S. Hall, G. T. Desmoulin, and T. E. Milner, “A technique for conditioning and calibrating force-sensing resistors for repeatable and reliable measurement of compressive force.,” Journal of biomechanics, vol. 41, no. 16, pp. 3492-3495, Dec. 2008. [53] J. Z. Zheng, S. D. L. Rosa, and A. M. Dollar, “An Investigation of Grasp Type and Frequency in Daily Household and Machine Shop Tasks,” in Proceedings of the International Conference on Robotics and Automation, 2011, pp. 4169-4175. [54] J. R. Flanagan and M. a Beltzner, “Independence of perceptual and sensorimotor predictions in the size-weight illusion,” Nature neuroscience, vol. 3, no. 7, pp. 737741, Jul. 2000. [55] G. Buckingham, N. S. Ranger, and M. A. Goodale, “The role of vision in detecting and correcting fingertip force errors during object lifting,” Journal of Vision, vol. 11, pp. 1-14, 2011. [56] E. C. Chang, J. R. Flanagan, and M. A. Goodale, “The intermanual transfer of anticipatory force control in precision grip lifting is not influenced by the perception of weight.,” Experimental Brain Research, vol. 185, no. 2, pp. 319-329, Mar. 2008. [57] H. Forssberg, A. C. Eliasson, H. Kinoshita, G. Westling, and R. S. Johansson, “Development of human precision grip. IV. Tactile adaptation of isometric finger forces to the frictional condition.,” Experimental brain research, vol. 104, no. 2, pp. 323-330, Jan. 1995. [58] J. M. Romano, S. Member, and K. Hsiao, “Human-Inspired Robotic Grasp Control With Tactile Sensing,” Transactions on Robotics, vol. 27, no. 6, pp. 1067-1079, 2011. [59] J. H. McDonald, Handbook of Biological Statistics, 2nd ed. Baltimore, Maryland: Sparky House Publishing, 2002.  90  [60] A. H. Mason and B. J. Bernardin, “The Role of Visual Feedback when Grasping and Transferring Objects in a Virtual Environment,” in Proceedings of the International Conference on Enactive Interfaces, 2008, pp. 111-116. [61] A. H. Mason and C. L. Mackenzie, “The Role of Graphical Feedback About SelfMovement when Receiving Objects in an,” Presence, vol. 13, no. 5, pp. 507-519, 2004. [62] A. Moon, “What Should a Robot Do? Design and Implementation of Human-like Hesitation Gestures as a Response Mechanism for Human-Robot Resource Conflicts,” M.A.Sc. Thesis, University of British Columbia, 2012. [63] J. Kim, J. Park, Y. K. Hwang, and M. Lee, “Advanced Grasp Planning for Handover Operation Between Human and Robot: Three Handover Methods in Esteem Etiquettes Using Dual Arms and Hands of Home-Service Robot Grasp Planners for Handover Operations,” in Proceedings of the International Conference on Autonomous Robots and Agents, 2004, pp. 34-39. [64] M. L. Walters, K. Dautenhahn, S. N. Woods, and K. L. Koay, “Robotic Etiquette: Results from User Studies Involving a Fetch and Carry Task,” in Proceedings of the International Conference on Human-Robot Interaction, 2007, pp. 317-324. [65] B. Zhu, “Design of Etiquette for Patient Robot Interaction in a Medicine Delivery Task,” M.Sc. Thesis, North Carolina State University, 2009.  91  Appendix A. Personal Communications A1. Correspondence with Dr. Travis Deyle Regarding Their Handover Controller Date: Fri, 27 Jan 2012 11:28:31 -0500 Subject: Re: Elder care robot project Hey Wesley, As Tiffany pointed out, the old code from EL-E probably isn't very useful. By and large, it was a home-grown software system that didn't run on ROS and honestly, it's been so long (3 years?) that I don't even know where that code would reside. But I can tell you everything I remember.... From a hardware perspective, there was a ATI 6-axis force-torque sensors at the base of each of ELE's fingers. For my RFID work, the robot would hold the medication bottle out to a fixed position relative to the robot (after checking to ensure no collisions, of course). From that point, it would start monitoring the FT sensors for changes (deltas in force / torque). I recall manually tuning a force threshold and a torque threshold -- if either was exceeded, the robot would release the object by opening its gripper. The thresholds were set so that drift in the sensors would not artificially trigger object release; they required a user to slightly tug on the object / fingers. I don't recall the exact threshold values (in Newtons / Newton-Meters), but they are probably pretty contingent on the mounting of the sensors (ie. longer fingers generate more torque). If the robot did not receive any delta-FT after a certain window (5 seconds, I believe), it would re-tuck it's arm and move on to the next action (ie. "handoff failure"). There is one obvious caveat to my approach with the RFID medication delivery: the robot would just hold out the meds to some fixed position relative to the robot. This is pretty rudimentary, since it wasn't the focus of my research. I seem to recall having conversations about being more clever about where to hold out the robot's arm (perhaps this was discussed in Young Sang Choi's "Hand It Over or Set It Down...", but I wasn't involved in that work). I think there is definitely room for improvement here -- so I look forward to seeing the results of your work! Now... the PR2 medication delivery used the same approach as EL-E, except that we used the Pressure Profile Systems (PPS) tactile sensors in the PR2's gripper instead of 6-axis force torque sensors. This has two drawbacks: (1) The PPS sensors only sense normal forces, so tugging on the object didn't work as well as "torquing" it, and (2) The PPS sensors' outputs are not calibrated in physical units (ie. Newtons). Thus... any attempt to tell you a threshold is rather arbitrary (what would "20" PPS-units mean?). That said... all of the code for this is open source and well-tested on ROS Diamondback! The major ones you (might) want to look at*: gt-ros-pkg/hrl/pr2_rfid/rfid_behaviors/src/rfid_behaviors/handoff_node.py gt-ros-pkg/hrl/pr2_rfid/rfid_behaviors/src/rfid_behaviors/tactile_sensors.py * I don't claim to be the best software engineer. ;-) Also, I've been told by labmates that some of the dependencies (specifically hrl_lib) might not function on newer versions of ROS (for whatever reasons).  92  Anyway, best of luck. Let me know if you have any other questions. Citations of the Pervasive paper (ie. for medication delivery) are also always welcome! Just be gentle on the critique. :-) Oh! And you should check out Hizook.com too! (Shameless plug.) Cheers, PS -- Just in case, I'm going to send this email from my (permanent) Gmail address. Should probably use the gmail address in the future since I'm no longer affiliated with Georgia Tech.  ~Travis Deyle NSF "Computing Innovation" (CI) Postdoc Fellow @ Duke University PhD from Georgia Tech's Healthcare Robotics Lab Founder www.Hizook.com -- Robotics News for Academics & Professionals  On Wed, Jan 25, 2012 at 3:53 PM, Wesley Chan wrote: Hi Dr Deyle, I am a student at the CARIS lab working on human-robot handovers. I was wondering if the code you used for PR2 med delivery (mentioned by Tiffany below) has been released or if it has been planned for release? In your paper "RFID-Guided Robots for Pervasive Automation" you talked about EL-E handing over objects to users and EL-E releases the object by sensing forces and torques above a threshold. What were the force torque threshold used and in what axis were they measured? Where can I find more information about that project? Thanks, Wesley Chan M.A.Sc. student Collaborative Advanced Robotics and Intelligent Systems (CARIS) Laboratory University of British Columbia  Date: Wed, 25 Jan 2012 12:45:21 -0500 Subject: Re: Elder care robot project Hi Wesley, I do not think the code written for EL-E (the Young Sang study) has been discussed for release. EL-E code was not written using ROS and no one uses that robot anymore so I imagine we will never actually release it. You can look at ROS code written by Travis Deyle on RFID who wrote the med delivery code for the PR2. It may or may not be released now. Tiffany  93  On Wed, Jan 25, 2012 at 12:25 PM, Wesley Chan wrote: Hi Tiffany, I was wondering if the source code used for the studies with older adults receiving medication bottles, or the source code used for Young-Sang's study is available for download? Or will they be made available some time later? Thanks, Wesley Date: Sun, 15 Jan 2012 16:19:56 -0500 Subject: Re: Elder care robot project Hello Wesley, I received your email below to my advisor Dr. Kemp and I wanted to point you at some papers that may be relevant. Sorry for the delayed response! You can view all of our lab's publications here: http://www.hsi.gatech.edu/hrl/publications.shtml You might be interested in the papers authored by Young-Sang Choi, particularly this one: http://www.hsi.gatech.edu/hrl/pdf/roman2009_delivery.pdf We haven't yet published the results from our studies with older adults receiving a medication bottle from our PR2, but check back for updates! We may publish a tech report in the spring. Hope that helps, Tiffany  94  ---------- Forwarded message ---------From: Wesley Chan Date: Wed, Jan 11, 2012 at 5:42 PM Subject: Elder care robot project Hi Dr. Kemp, I am a masters student of Mike Van der Loos working on human-robot handovers. Mike said that some time ago when he got a tour of your lab he was shown a video of an elder care robot that dispenses medication for patients. I am interested in finding out more about the project and am wondering if there are any related papers or materials you can point me to. Thanks, Wesley Chan M.A.Sc. student Collaborative Advanced Robotics and Intelligent Systems (CARIS) Laboratory University of British Columbia -Tiffany L. Chen Georgia Institute of Technology Department of Biomedical Engineering Healthcare Robotics Laboratory www.healthcare-robotics.com http://www.prism.gatech.edu/~tchen46/index.html  95  A2. Correspondence with Dr. Andrea Mason Regarding Baton Size Design Date: Wed, 8 Jun 2011 13:33:37 -0500 Subject: Re: research on grip force and haptic interaction in handovers Hi Wesley, Sorry for the delay in replying but I've been on vacation for the past several days. I haven't done a whole lot in the way of passing research since my time at SFU, but I've attached a few papers that we did looking at passing coordination in a virtual environment. Note that we didn't measure grip forces in these papers. In general, when choosing an object size for precision grasping, we try to choose an object that is "comfortable to grasp" over repeated trials. If the object is too wide, the hand will become fatigued. A width of less than 2.5" is preferable. You want the length of the object to be such that the two subjects can easily grasp without touching (i.e. 3" or so). Often we are limited when choosing the size of the object based on what measurement systems we are using (i.e. will there be transducers on load cells embedded, do LEDs for motion tracking have to be placed on the object, etc.). Please don't hesitate to contact me again if you have further questions. Dr. Mason On 6/2/2011 12:27 PM, Wesley Chan wrote: Dear Dr Mason, Dr MacKenzie, I am a master student working in the Collaborative Advanced Robotics and Intelligent Systems (CARIS) Laboratory at the University of British Columbia under Elizabeth Croft and Mike Van der Loos. I am researching in the haptic aspect of human-robot handovers and am very interested in some of your work. I've read your paper "Grip forces when passing an object to a partner" and have been reading Dr MacKenzie's book "The Grasping Hand". I was wondering if you have any other papers or publications on handovers that you can point me to. Also, I will be conducting a study in human-human handover and am wondering if there are any guidelines to choosing the size of the object for handing over. How did you choose the size of the object used in your experiment in the paper "Grip forces when passing an object to a partner"? Thanks, Wesley Chan M.A.Sc. candidate Collaborative Advanced Robotics and Intelligent Systems (CARIS) Laboratory University of British Columbia -Andrea Mason Associate Professor, Kinesiology  _________________________________ Phone ___________  96  Appendix B. Human-Human Object Handover Study Materials B1. Participant Consent Form The University of British Columbia Collaborative Advanced Robotics and Intelligent Systems (CARIS) Laboratory Department of Mechanical Engineering, UBC 6250 Applied Science Lane, Vancouver, BC V6T 1Z4 Tel: ___________ Fax: ___________ Web site: http://caris.mech.ubc.ca  HRI-Cues: Human-Human Handover Study Project Title: HRI-Cues: Human-Human Handover Study Principal Investigator: Elizabeth Croft Co-Investigator and Contact Person: Wesley Chan Funding: This research is funded by the Natural Sciences and Engineering Research Council of Canada (NSERC). Purpose: The purpose of this study is to characterize the haptic interactions in human-human handovers, and to explore how a human giver and receiver regulate their grip forces according to their experienced load forces during a handover. Results from this study will be used in subsequent research to improve the ability of robotic assistants to interact with non-expert human users. Procedures: Before the actual handover experiment, you will be asked to fill out a preliminary questionnaire which asks you for some demographic information (e.g., age, gender, etc). For the experiment, you will be paired with another participant to perform a series of handovers using a baton while wearing a motion sensor on your wrist. One participant will play the role of the giver, while the other will be the receiver. During the experiment, the giver and the receiver will be seated across a table. The giver will start with the baton in his/her hand, and both the giver and receiver will start with their hands placed over the table in front of them. On a verbal “Go” signal, the giver will move from the start position into an “interception zone” marked in the middle of the table. Once the giver has reached the interception zone, the receiver will move into the interception zone to take the object from the giver. After object transfer, both the giver and the receiver will return to their start positions. Three different baton weights will be used, and after a number of handovers the giver and the receiver will be asked to switch roles. The experiment will last approximately 45 minutes. You may refuse to participate in this experiment and you may withdraw at any time. We also would like to videotape this experiment, although this is not required for your participation. Potential Risks: Slight temporary fatigue from passing a baton back and forth.  97  Confidentiality: No identifying information will be collected or stored with your data. Data collected during the survey will be stored on a password protected computer in the CARIS Lab, which has restricted secure access and is locked at all times. If you have any concerns about your treatment or rights as a research participant, you may telephone the Research Subject Information Line in the UBC Office of Research Services at the University of British Columbia, at ___________. Consent: By signing this form, you consent to participate in this study, and acknowledge you have received a copy of this consent form. I agree to allow myself videotaped during this experiment (please circle one): yes no Name (print):______________________________________ Date:_________________ Signature:_______________________________________________  Version date: June 24, 2011  98  Appendix C. Robot-Human Object Handover Study Materials C1. Participant Consent Form The University of British Columbia Collaborative Advanced Robotics and Intelligent Systems (CARIS) Laboratory Department of Mechanical Engineering, UBC 6250 Applied Science Lane, Vancouver, BC V6T 1Z4 Tel: ___________ Fax: ___________ Web site: http://caris.mech.ubc.ca  HRI-Cues: Human-Robot Handover Study Project Title: HRI-Cues: Human-Robot Handover Study Principal Investigator: Elizabeth Croft Co-Investigator and Contact Person: Wesley Chan Funding: This research is funded by the Natural Sciences and Engineering Research Council of Canada (NSERC) and General Motors (GM). Purpose: The purpose of this study is to compare different robot handover controllers, and to determine user preferences among the controllers. Results from this study will be used in subsequent research to improve the ability of robotic assistants to interact with non-expert human users. Procedures: During the study, different handover controllers will be presented in pairs to you and you will be asked to rate the controllers. For each controller pair, you will be asked to do three handovers with one controller, followed by another three handovers with the second controller. After each pair, we will ask you to fill out a short survey to rate the two controllers. There will be a total of six pairs of controllers. The experiment will last for approximately 45 minutes. You may refuse to participate in this experiment and you may withdraw at any time. We also would like to videotape this experiment, although this is not required for your participation. Videos recorded will be used for analysis of the handover motion, and select video clips may be used in presentations, publications, and project related websites to illustrate the handover motion between the subject and the robot. Identifying features (e.g. faces) will be blurred out of any videos prior to their inclusion in publications/presentations/websites. Only the researchers involved with this study will be able to view the video recordings otherwise. Potential Risks: None. Confidentiality: No identifying information will be collected or stored with your data. Data collected from the survey will be stored on a password protected computer or a locked cabinet in the CARIS Lab, which has restricted secure access and is locked at all times. If you have any concerns about  99  your treatment or rights as a research participant, you may telephone the Research Subject Information Line in the UBC Office of Research Services at the University of British Columbia, at ____________. Consent: By signing this form, you consent to participate in this study, and acknowledge you have received a copy of this consent form. I agree to allow myself videotaped during this experiment (please circle one): yes no Name (print):______________________________________ Date:_________________ Signature:_______________________________________________  Version date: July 10, 2012  100  C2. Experiment Survey Experiment Number: _______ Rate the likelihood that the object could have been dropped during the handover. Likely to be dropped 1 2 3 4 5 6 7 Not likely to be dropped Controller 1: Controller 2:  □ □  □ □  □ □  □ □  □ □  Rate how easy it was to take the object in the handover. Very hard to take 1 2 3 4 5  7  □ □  □ □  Rate how much you prefer each handover. Not at all preferred 1 2 3 4 preferred  5  6  7  □ □  □ □  □ □  □ □  Controller 1: Controller 2:  □ □  □ □  6  □ □  Controller 2:  □ □  □ □  □ □  Controller 1:  □ □  □ □  □ □  □ □  Very easy to take  Very much  Please provide any comments you have:  __________________________________________________________ __________________________________________________________ __________________________________________________________ __________________________________________________________ 101  __________________________________________________________ __________________________________________________________ Version date: July 4, 2012  102  Appendix D. Wilcoxon Signed-Rank Test Results for RobotHuman Handover Study Survey Responses Table D113. Analysis results for comparing likelihood of dropping the object. Significant results are marked with **.  Controller A vs. Controller B Controller A vs. Controller C Controller A vs. Controller D Controller B vs. Controller C  Negative Positive Differences Differences 11 19 14 19 34 2 7 7  Ties 18 15 12 34  Wilcoxon Signed-Rank Test Result Z = -1.437, p = 0.151 Z = -0.331, p = 0.740 Z = -4.942, p < 0.0005** Z = -0.233, p = 0.816  Table D2. Analysis results for comparing ease of taking the object. Significant results are marked with **.  Controller A vs. Controller B Controller A vs. Controller C Controller A vs. Controller D Controller B vs. Controller C  Negative Positive Differences Differences 42 1 44 1 5 30 25 14  Ties 5 3 13 9  Wilcoxon Signed-Rank Test Result Z = -5.724, p < 0.0005** Z = -5.843, p < 0.0005** Z = 3.982, p < 0.0005** Z = -1.806, p = 0.071  Table D3. Analysis results for comparing user preference. Significant results are marked with **.  Controller A vs. Controller B Controller A vs. Controller C Controller A vs. Controller D Controller B vs. Controller C  Negative Positive Differences Differences 41 3 41 2 29 11 25 11  Ties 4 5 8 12  Wilcoxon Signed-Rank Test Result Z = -5.234, p < 0.0005** Z = -5.505, p < 0.0005** Z = -3.445, p = 0.001** Z = -2.112, p = 0.035**  103  

Cite

Citation Scheme:

        

Citations by CSL (citeproc-js)

Usage Statistics

Share

Embed

Customize your widget with the following options, then copy and paste the code below into the HTML of your page to embed this item in your website.
                        
                            <div id="ubcOpenCollectionsWidgetDisplay">
                            <script id="ubcOpenCollectionsWidget"
                            src="{[{embed.src}]}"
                            data-item="{[{embed.item}]}"
                            data-collection="{[{embed.collection}]}"
                            data-metadata="{[{embed.showMetadata}]}"
                            data-width="{[{embed.width}]}"
                            async >
                            </script>
                            </div>
                        
                    
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:
http://iiif.library.ubc.ca/presentation/dsp.24.1-0073332/manifest

Comment

Related Items