UBC Theses and Dissertations

UBC Theses Logo

UBC Theses and Dissertations

Exploring structured predictions from sensorimotor data during non-prehensile manipulation using both… Zhu, Chuan 2013

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

Item Metadata


24-ubc_2014_spring_zhu_chuan.pdf [ 12.8MB ]
JSON: 24-1.0165671.json
JSON-LD: 24-1.0165671-ld.json
RDF/XML (Pretty): 24-1.0165671-rdf.xml
RDF/JSON: 24-1.0165671-rdf.json
Turtle: 24-1.0165671-turtle.txt
N-Triples: 24-1.0165671-rdf-ntriples.txt
Original Record: 24-1.0165671-source.json
Full Text

Full Text

Exploring Structured Predictions from Sensorimotor Dataduring Non-prehensile Manipulation using bothSimulations and RobotsbyChuan ZhuB. S. in Computer Science, Zhejiang University, 2011A THESIS SUBMITTED IN PARTIAL FULFILLMENTOF THE REQUIREMENTS FOR THE DEGREE OFMaster of ScienceinTHE FACULTY OF GRADUATE AND POSTDOCTORALSTUDIES(Computer Science)The University Of British Columbia(Vancouver)November 2013c? Chuan Zhu, 2013AbstractRobots are equipped with an increasingly wide array of sensors in order to en-able advanced sensorimotor capabilities. However, the efficient exploitation of theresulting data streams remains an open problem. We present a framework for learn-ing when and where to attend in a sensorimotor stream in order to estimate specifictask properties, such as the mass of an object. We also identify the qualitativesimilarity of this ability between simulation and robotic system. The framework isevaluated for a non-prehensile ?topple-and-slide? task, where the data from a setof sensorimotor streams are used to predict the task property, such as object mass,friction coefficient, and compliance of the block being manipulated. Given the col-lected data streams for situations where the block properties are known, the methodcombines the use of variance-based feature selection and partial least-squares es-timation in order to build a robust predictive model for the block properties. Thismodel can then be used to make accurate predictions during a new manipulation.We demonstrate results for both simulation and robotic system using up to 110sensorimotor data streams, which include joint torques, wrist forces/torques, andtactile information. The results show that task properties such as object mass,friction coefficient and compliance can be estimated with good accuracy from thesensorimotor streams observed during a manipulation.iiPrefaceThe motivation section of Chapter 1 is inspired by the research challenges docu-mented by Kemp, Edsinger, and Torres-Jara [KETJ07]. We build on their sum-maries and classification schemes.The description of prediction approaches given in Chapter 3 is loosely basedon the overview of this class of problem provided in [HTF01]. We appreciate theirtutorial and introduction to modern machine learning approaches.Chapter 4, 5 and 6 are based on work conducted in UBC?s IMAGER lab byDaniel Troniak, Chuan Zhu, Dinesh K. Pai, Michiel van de Panne. I was respon-sible for physically-based simulation, simulation data collection,data analysis forboth robots and simulations, experiments implementation using matlab. Danielmainly worked on robot control and maintenance, and robot data collection. Thedesign of experiments, result generation and algorithm were developed in collabo-ration. A version of this work without the simulation results [TZPvdP14] has beensubmitted to ICRA?2014. Several figures and part of texts from the publication arecopyright and are reused in this thesis by permission.iiiTable of ContentsAbstract . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiPreface . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . iiiTable of Contents . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . ivList of Tables . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiList of Figures . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . viiiAcknowledgments . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . xi1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1 Motivation . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 11.1.1 Simulation . . . . . . . . . . . . . . . . . . . . . . . . . 11.1.2 Controlled Environments . . . . . . . . . . . . . . . . . . 21.1.3 Human-populated (Uncontrolled) Environments . . . . . . 31.1.4 Our Motivation . . . . . . . . . . . . . . . . . . . . . . . 61.2 Challenges . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 71.3 Goals . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 81.4 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . . . 91.5 Thesis Organization . . . . . . . . . . . . . . . . . . . . . . . . . 92 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 102.1 Perception in Robotics . . . . . . . . . . . . . . . . . . . . . . . 102.1.1 Vision-based Sensing . . . . . . . . . . . . . . . . . . . . 10iv2.1.2 Tactile Sensing . . . . . . . . . . . . . . . . . . . . . . . 112.2 Learning . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122.3 Non-prehensile Manipulation . . . . . . . . . . . . . . . . . . . . 132.4 Dimensionality Reduction . . . . . . . . . . . . . . . . . . . . . 142.4.1 Feature Extraction . . . . . . . . . . . . . . . . . . . . . 142.4.2 Feature Selection . . . . . . . . . . . . . . . . . . . . . . 152.5 Anomaly Detection in Streaming Data . . . . . . . . . . . . . . . 153 The Block Topple-and-Slide Task . . . . . . . . . . . . . . . . . . . . 164 Prediction Approaches . . . . . . . . . . . . . . . . . . . . . . . . . . 204.1 Probabilistic Classification Model . . . . . . . . . . . . . . . . . 204.2 Linear Regression . . . . . . . . . . . . . . . . . . . . . . . . . . 214.2.1 Least Squares . . . . . . . . . . . . . . . . . . . . . . . . 224.2.2 Principal Components Regression . . . . . . . . . . . . . 234.2.3 Partial Least Squares . . . . . . . . . . . . . . . . . . . . 245 Structured Prediction Methodologies . . . . . . . . . . . . . . . . . 275.1 System Overview . . . . . . . . . . . . . . . . . . . . . . . . . . 275.2 Feature Selection via the Task Variance Ratio . . . . . . . . . . . 285.3 Property Estimation with Partial Least Squares . . . . . . . . . . 295.4 Online Estimation . . . . . . . . . . . . . . . . . . . . . . . . . . 306 Experiments and Results . . . . . . . . . . . . . . . . . . . . . . . . 326.1 Robotic Configuration and Experimental Testbed . . . . . . . . . 326.2 Simulation Configuration and Experimental Testbed . . . . . . . . 366.3 Procedure . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 376.3.1 Obtain the Task Trajectory . . . . . . . . . . . . . . . . . 376.3.2 Obtain the Sensorimotor Training Data, D . . . . . . . . . 376.3.3 Feature Selection . . . . . . . . . . . . . . . . . . . . . . 386.3.4 Partial Least Squares Modeling . . . . . . . . . . . . . . 386.3.5 Online Estimation . . . . . . . . . . . . . . . . . . . . . 386.4 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 386.4.1 Feature Selection Performance . . . . . . . . . . . . . . . 39v6.4.2 Beta Reduction Performance . . . . . . . . . . . . . . . . 426.4.3 Validation of PLS . . . . . . . . . . . . . . . . . . . . . . 426.4.4 Online Property Estimation . . . . . . . . . . . . . . . . 446.4.5 Qualitative Similarity . . . . . . . . . . . . . . . . . . . . 456.5 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 487 Conclusions and Future Work . . . . . . . . . . . . . . . . . . . . . 49Bibliography . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 51A Supporting Materials . . . . . . . . . . . . . . . . . . . . . . . . . . 60viList of TablesTable 4.1 Comparisons of different prediction approaches . . . . . . . . 26Table 6.1 Task property for robot with hand. . . . . . . . . . . . . . . . 33Table 6.2 Task property for robot with spherical probe. . . . . . . . . . . 33Table 6.3 Task property for simulation . . . . . . . . . . . . . . . . . . . 36Table 6.4 Task parameters . . . . . . . . . . . . . . . . . . . . . . . . . 37Table 6.5 Figure illustration . . . . . . . . . . . . . . . . . . . . . . . . 38viiList of FiguresFigure 1.1 OpenRAVE platform [Dia10] drives a simulated WAM robotarm to perform tasks such as carrying a cup and opening acabinet door. . . . . . . . . . . . . . . . . . . . . . . . . . . 2Figure 1.2 Automated assembly of crankshafts using a Comau robot atHams Hall engine plant (figure from [Mor07]). . . . . . . . . 3Figure 1.3 Work at Stanford lab with their STAIR-1 robot (figure from[KSN10]). (a)Autonomous object recognition. (b)Openingdifferent types of doors. . . . . . . . . . . . . . . . . . . . . 5Figure 1.4 Human Cortical Homunculus (figure from [Tat13]). The dis-torted appearance of the homunculus is due to the amount ofcerebral tissue or cortex devoted to a given body region, whichis proportional to how richly innervated that region is, not to itssize. The large hand indicates the significance of it to humansensing. . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 7Figure 3.1 The topple-and-slide task. The first row is the conceptual de-piction of a clear toppling and sliding action. The second rowshows toppling and sliding for the robot with hand. The thirdrow is this action for the robot with spherical probe. The lastrow shows a physically-based simulation of the task. . . . . . 17Figure 3.2 Visualization for the sensorimotor trace, x. Each uniquely col-ored segment represents one sensory sample s. The wholetrace consists of the observations of ns sensors each sampledat nt points in time. . . . . . . . . . . . . . . . . . . . . . . . 18viiiFigure 3.3 Visualization of the sensorimotor data stream D. Each patchwith a unique color represents the sensory sample s at one timephase. The full horizontal dimension represents one sensori-motor trace x for one property variation with n sensor read-ings, which is stacked into a single vector with ns sensors eachsampled at nt points in time. The vertical dimension representsall np variations of property set. Lastly, the depth dimensionrepresents all nh trials on one same task property variation. . . 19Figure 4.1 Linear Least Square fitting with X ? R2 [HTF01]. . . . . . . . 22Figure 4.2 Principal components of X ? R2 [HTF01]. The largest prin-cipal component is along the direction which maximizes thevariance of the projected data. . . . . . . . . . . . . . . . . . 24Figure 4.3 Schematic outline of PLS regression. The extracted input vari-ables X S (also referred to as predictor scores) are used to pre-dict extracted response scores y S and then the predicted y Sare used to construct predictions for the response. The coeffi-cient matrix ? can also be derived via this process. . . . . . . 25Figure 4.4 Prediction approach hierarchy . . . . . . . . . . . . . . . . . 26Figure 5.1 System Overview . . . . . . . . . . . . . . . . . . . . . . . . 28Figure 6.1 Testing friction coefficient using an inclined plane . . . . . . 34Figure 6.2 Property setup for robots . . . . . . . . . . . . . . . . . . . . 35Figure 6.3 Simulation environment . . . . . . . . . . . . . . . . . . . . 36Figure 6.4 Visualization of features selected according to the task vari-ance ratio, ?, of joint position, velocity, torque, and Cartesianpose measurements. . . . . . . . . . . . . . . . . . . . . . . 40Figure 6.5 Effect of TVR feature selection on mass estimation . . . . . . 41Figure 6.6 Effect of partial least squares dimensionality reduction on massestimation. . . . . . . . . . . . . . . . . . . . . . . . . . . . 42Figure 6.7 LOOCV for mass estimation on NBC, LS, PCR and PLS. . . 43ixFigure 6.8 Online estimation of the mass from sensorimotor data. Theestimates of mass are shown at various blue points throughoutthe motion. The dotted red horizontal line denotes the actualmass of the block. . . . . . . . . . . . . . . . . . . . . . . . . 44Figure 6.9 Online estimation of the friction coefficient from sensorimotordata. The estimates of friction coefficient are shown at variousblue points throughout the motion. And the dotted red hori-zontal line denotes the actual friction coefficient. . . . . . . . 45Figure 6.10 Qualitative similarity in task variance ratio, ? . . . . . . . . . 47xAcknowledgmentsI would never have been able to finish my thesis without the guidance of my com-mittee members, help from friends, and support from my family and my love.I would like to express my deepest gratitude to my supervisor, Michiel van dePanne, for his excellent guidance, caring, patience, financially support and provid-ing me with an excellent atmosphere for doing research. I would also like to thankDinesh K. Pai for guiding my research for the past year. Special thanks goes toDaniel Troniak, who as a best friend and teammate, makes so many efforts on ourproject. Many thanks to Shuo Shen, Jingxian Li, Shailen Agrawal, Boris Dalstein,David Brown and all other workers in the laboratory for helping me. I would alsolike to thank my parents. They were always supporting me and encouraging mewith their best wishes. Finally, I would like to thank my love, Xi Chen. She wasalways there cheering me up and stood by me through the good times and bad.xiChapter 1Introduction1.1 MotivationRobots are widely used in many applications today, including surgery, health careand natural product processing. Robots may be the feasible substitutes for manyforms of manual work in the future due to their flexibility, reliability and ability todo boring or dangerous tasks. Robots can perform many tasks that people can do aswell as being qualified to perform other tasks not well suited to humans such as res-cue excavation from earthquake rubble. With ongoing advances in manufacturingand technology, robots are becoming increasingly skillful and adept.Robotics research can be classified into three categories [KETJ07]: simulation,controlled environments and human environments. Here, we also broadly followthese categories. Robots have been very successful at manipulation in simulationand controlled environments but are challenged by uncontrolled environments dueto uncertainty and possible perturbation.1.1.1 SimulationA robotics simulator is used to create embedded applications for a robot withoutdepending physically on the actual machine. Simulated robots have accomplishedvarious manipulation tasks such as grasping cubes, carrying objects, and openinga door (Figure 1.1). Control policies for these demonstrations often employ searchalgorithms to find satisfactory solutions, such as depicting the optimal path when1given a desired goal [KETJ07]. Given that the physical rules are fully known andthat noise does not exist in simulated environments, such a system is highly ide-alized. Most simulations neglect the robot?s sensory systems due to constraints ofthe simulator and the assumption that the world geometry is fully known. If devel-opers use a robot model that is not close enough to the actual robot dynamics andcapabilities, then results can be meaningless [PBS07]. As a result, a robotics sim-ulator might only be useful in the initial phases of development and the resultingmotions or skills could be hard to transfer to a real robot.Figure 1.1: OpenRAVE platform [Dia10] drives a simulated WAM robot armto perform tasks such as carrying a cup and opening a cabinet door.1.1.2 Controlled EnvironmentsFor controlled environments, the world can be arranged to accommodate the knownabilities of the robot. For example, on an automated industrial assembly line (Fig-ure 1.2), engineers can ensure that a robot knows the relevant state of the world withnear certainty and make the configuration of the assembly line conform to what isrequired by the robots. In such a scenario, it is common to forbid people fromentering the work area. Researchers can also simplify the environment settingswithin which their robot operates in order to only focus on the specified problems2of interest. In addition, in the event that a robot needs to sense the world, engineerscan make the environment favorable to sensing by controlling factors such as thestrength of the lighting and the placement of objects relative to a sensor [KETJ07].We quote the conclusion in [KETJ07] that thus far, ?successful demonstrationsof research robots autonomously performing complicated manipulation tasks haverelied on some combination of known objects, simplified objects, uncluttered en-vironments, fiducial markers, or narrowly defined, task-specific controllers.?Figure 1.2: Automated assembly of crankshafts using a Comau robot atHams Hall engine plant (figure from [Mor07]).1.1.3 Human-populated (Uncontrolled) EnvironmentsUnlike in well controlled environments, natural settings in the real world cannot beperfectly modeled. The following list summarizes some of the challenging charac-teristics of common uncontrolled environments:3? Human intervention ? Users who are not roboticists may be nearby.? Environment complexity ? A natural human environment is not as simplifiedas the setting in the lab.? Non-reproducibility ? The same environment configuration is impossible tobe reproduced in any two different locations.? Dynamic change ? The world configuration is easily changed at any time.? Lack of monitoring ? No one can keep an eye on the robot movement all thetime.? Perturbation variation ? The real human environment is full of unpredictableeffects, such as lighting variation, occluding objects and background sounds.Although people are able to handle these issues successfully without any assis-tance, the lack of facility to cope with such dramatic variations impacts heavily ona robots? behavior. For example, if you were to visit a home that you have nevervisited before, you would have no trouble to use the bathroom even if everything isquite different from your own house. By contrast, although it is not hard for robotsto find where is the door of the bathroom by using a visual system (Figure 1.3),they would likely have difficulties with autonomously using a novel type of faucet.The reliable introduction of robots into human environments will highly dependon the development of motion generation, perception, knowledge management anddecision making technologies [SYKM06].4(a)(b)Figure 1.3: Work at Stanford lab with their STAIR-1 robot (figure from[KSN10]). (a)Autonomous object recognition. (b)Opening differenttypes of doors.51.1.4 Our MotivationIt is a challenge to make robots execute tasks robustly in unstructured, human-centered environments. Since it is infeasible to explicitly model every possiblescenario within this context, robots must be able to learn through the experience ofinteracting with its dynamic environment [SPK02]. For the many cases where vi-sual systems cannot provide complete information, we believe that the robot shouldbe able to estimate task-relevant properties, such as the mass of an object, its fric-tion coefficient, or its compliance, indirectly from other available sensory informa-tion, such as joint torques, end-effector forces/torques, or tactile pressure readings.For example, when executing a sliding task, the heavier the box is, the larger theforce on finger tips and on the wrist will be. As shown in Figure 1.4, we candemonstrate how important sensory information is to humans for the hands. If weexamine the significant resources allocated to sensing from the hands, it is clear thatthey serve an important role. By learning to cope with changing dynamics, a robotstands a better chance of succeeding in its task despite encountering environmentconfigurations with which it may be unfamiliar.6Figure 1.4: Human Cortical Homunculus (figure from [Tat13]). The distortedappearance of the homunculus is due to the amount of cerebral tissue orcortex devoted to a given body region, which is proportional to howrichly innervated that region is, not to its size. The large hand indicatesthe significance of it to human sensing.1.2 ChallengesMost environment properties, e.g. the mass of an enclosed coffee cup, cannot bedirectly observed without specialized tools and can therefore only be estimated byinteracting with the object. The problem of estimating physical properties in awell-controlled environment using specially designed exploratory procedures hasbeen previously addressed. However, for a changing and uncertain environment,this problem has not yet been well explored.The available sensory data streams are typically high dimensional, which makesthe problem of using these sensors a challenge one. For example, a set of 100 sen-sors sampled at 100 Hz and collected during a prescribed 10 s manipulation task7provides 105 input features that can be used to estimate the desired properties.Learning from high-dimensional data, such as data collected during a robotic ma-nipulation, remains an open problem [Neu13].A related challenge is to make online estimates using only the data gatheredto date. If particular properties change during the task, e.g., a cup is lighter thanexpected or a block slides less easily than expected, it would be useful for the robotto identify this during online operation and to adapt its actions accordingly.1.3 GoalsOur proposed approach to predict from sensorimotor streams is model-free [SC98]in that there is no kinematic or dynamic model that provides an apriori predictiverelationship between the motion, the sensorimotor data, and the task properties. Inour setting, the associations between the motion, the sensor readings, and the taskproperties must be learned from example data. Non-prehensile manipulation arean ideal application for this type of model-free approach to the robot control. Innon-prehensile manipulation, a manipulator does not have complete control overan object being manipulated and important properties, such as sliding friction, canchange rapidly.The first goal of our approach is to automatically identify particular sensors atparticular phases of the motion that serve as the most important features for theproperty prediction. This attentional approach means that only a small fraction ofthe sensory data needs to be streamed off the robot platform.Our second goal is to be able to estimate the unknown environment property atany point during the non-prehensile manipulation. These estimates can be updatedas additional sensory information streams in over time.Lastly, we are also interested in finding correspondences between simulationsand robots operating in the real-world. Exploring and experimenting on the robotsin the real-world is not feasible all the time due to the expense and risks of workingwith hardware. Hence, it would be useful to be able to demonstrate a qualitative orquantitative connection between simulation results and expected real-world results.81.4 ContributionsThis thesis investigates an alternative approach that enables the reliable estimationof mass, friction coefficient, and compliance properties from a stream of unlabeledsensory data while performing a prescribed manipulation task in real time. Thepredictive model is developed with the help of training data, where the same ma-nipulation task is first carried out multiple times on objects with known properties.Analysis of the sensory data collected during these trials then identifies which sen-sory readings, i.e., which sensors at which time steps, are most likely to provideuseful information for the estimation problem. Based on these selected features, apartial least squares regression model is then developed to make the desired prop-erty predictions. We demonstrate that our approach can provide a qualitativelysimilar results on both the simulation platform and the real robot application.1.5 Thesis OrganizationIn Chapter 2, we review previous work relating to perception, learning and di-mensionality reduction methodologies used in robotics. We also briefly reviewmethods for the control of non-prehensile manipulation. In Chapter 3, we describethe non-prehensile manipulation task that is the primary focus of this thesis, whichis a ?topple-and-slide?task. Chapter 4 introduces and focuses on the learning ap-proaches which we will take into consideration in this paper. In Chapter 5, wedocument our structured prediction system and present the detailed methodologiesto implement each procedure in our framework. The experiments and results aredescribed in Chapter 6 where we also demonstrate and analyze the performance ofour method. We conclude in Chapter 7 with a summary of contributions, limita-tions and also lay out directions for future work.9Chapter 2Related Work2.1 Perception in RoboticsIn unstructured environments, robots will often need to cope with uncertainty. Thisintroduces the need for perception, including vision, touch, force and other possiblesensing modalities.2.1.1 Vision-based SensingVision and manipulation, which are often inextricably intertwined, have been stud-ied for many decades. In what follows below, we only discuss a very small sub-set of the large volume of research that is relevant to the use of computer visionin robotics. Localizing robot itself accurately from visual landmarks is first pro-posed in [AH93] by using a single geometric tolerance to describe observationerror. The authors in [SLL01] present a vision-based mobile robot localization andmapping algorithm using scale-invariant image features (SIFT features, [Low99]and [Low04]) as visual landmarks and extend it to a global scenario later [SLL05].The grasping issue could be addressed by marking the locations of the grasp-ing points and storing the relevant features for the assessment of grasp quality[KFE96]. The problem of manipulation of novel objects or in new environmentshas become a problem of considerable interest in recent years. In [SDK+06], alearning-based approach enables successfully grasping novel objects without rely-ing on a 3-d model. Using a monocular camera, they detect the best position at10which to grasp the object. Similarly, the problem of opening new doors under theassumption of no prior knowledge of location of the door and shape of the handle isalso well addressed using object recognition [KSN10]. This work is an example ofa simulation-based and learning-based approach that is successfully transferred to arobot operating in the real-world. However, vision-based approaches in robotic ma-nipulation are often limited in their use outside of laboratory and industrial setting[SHC92], in part because it typically has specific requirements related to suitablelighting.2.1.2 Tactile SensingTactile sensing is another more intuitive and general approach for guiding manipu-lation. Human dexterity stems in part from the physical structure of primate hands,especially the ability to sense conditions at the finger-object contact [How93].Hence, tactile sensing seems like an especially appropriate modality to improvethe reliability and feasibility of robotics manipulation. As anyone working in thedark can convincingly demonstrate, using tactile information alone is sufficient tocomplete many complicated tasks [KETJ07]. Another experiment, performed byanesthetizing the skin on the hands of a group of volunteers, also demonstratesthe significant role of the sense of touch for maintaining a stable grasp of object[WJ84].A detailed survey on the tactile sensing modality has recently been presentedin [DMVS10]. Tactile sensing is a form of sensing that can establish many use-ful properties of an object, such as finding contact locations, perceiving its shape,measuring contact forces, and determining contact conditions [How93], [Lee00].Manipulation based on tactile sensing allows a robot to quickly and accurately de-tect changes in the touchable environment. As a result, a series of object properties,such as surface texture and object shape, can be sensed both statically [TLG73]and dynamically [Ell87], [GLP84], [HC93]. The transient event information, suchas contact events, local skin curvature and incipient slip, has been detected usinga multi-element stress rate sensor which consists of piezoelectric polymer stripsmolded into the surface of the rubber covering the robot finger tip [SMH94]. Theincipient slip detection is crucially based on the fact that at the moment when the11object starts to slip the rubber is stretched [Mel01]. In [Mel00], C. Melchiorri de-velops a model to detect rotational and translational slip using force/torque andmatrix tactile sensor. Tactile sensing also helps robots in adapting to their environ-ment. In [PRKS11], a robot learns the trajectory required by grasping an objectthrough kinesthetic teaching and records expected sensorimotor data streams to in-crease the likelihood of subsequent task success in the face of small environmentalperturbations. In our work, we focus on gaining deeper insight into the chang-ing environment so that the robot can succeed even when large perturbations areexperienced.2.2 LearningRobotic manipulation is most successful in carefully engineered environments wherethe physical characteristics of the world are well-understood and the effects ofrobot actions can be expected. For such situations, a manipulation procedure that ismanually developed by programmers can successfully cover the limited variations.But in the world which is poorly behaved such as the human-centered environment,we can often not reliably predict what will happen due to uncertainty and instabil-ity. In such cases, it is more profitable to rely on learning approaches in order toadapt to changes using the learned knowledge.As described above, both vision-based and tactile sensing approaches relyheavily on machine learning techniques in order to infer some unobservable or un-known properties of the world. In the 1990?s, many systems were developed thatfollow the Programming by Demonstration (PbD) paradigm [ADA89], [FHD98],[KH97], [KI97], focusing on the task of reconstructing goal-directed gestures ortrajectories despite different initial conditions and changes occurring during taskexecution [HGCB08]. The key purpose of robotic manipulation then becomesan interactive process of acquiring any change about the environment and sub-sequently adapting the interactions with the world in response to this change. In[KPB08], the robot autonomously learns manipulation strategies by physically in-teracting with the object and selects interactions which reveals the maximum in-formation about the kinematic structure when it deals with unknown articulatedobjects. Learning prehensile affordances through interaction with various manipu-12lable objects has been studied in [DKK+11]. The authors allow the robot to ?play?with a series of objects in order to learn and refine grasp affordances through explo-ration of the object pose space relative to the robot. In [FMN+03], the authors workon learning a model for the effects and consequences of self-generated actions, forexample, how the robot can pull an object toward itself or push it away. A learningframework that enables robots to acquire manipulation knowledge by observing theoutcomes of exploratory movements on objects is presented in [OBA+09]. Robotscan realize whether a grasping action will be successfully executed without theneed to first move the object to a better location.2.3 Non-prehensile ManipulationIn the current work, we are concerned with non-prehensile manipulation tasks, aspioneered by Lynch and Mason [Mas86], in which the robot does not have com-plete control over the object being manipulated. Nevertheless, non-prehensile ma-nipulation has seen relatively little attention in robotics literature despite occupyingthe majority of the manipulation spectrum: an important reason may be due to thefact that non-prehensile object manipulation is difficult to model explicitly [Lyn98].In [Lyn99], the authors take an analytical approach to the non-prehensile topplingtask. This approach, while successful, assumes knowledge of the dynamics of theentire system and would therefore not generalize to operation outside of controlledfactory environments where complete models of robot-object interaction dynamicswere unavailable. More recently, in [ZL12], the authors take a model-based ap-proach to the non-prehensile task of manipulating an object with rolling contactsacross a robotic finger-tip using tactile sensor feedback. While successful, this ap-proach relies heavily on accurate a priori kinematic and dynamic models of boththe environment and the robot.Another recent method for predicting the action outcomes of non-prehensilemanipulation is explored in [RUCB10]. Based on apriori knowledge of the massand friction coefficient, the authors design a simple but efficient approach to detectwhether the box is rotated or toppled when sliding on the table by only using theforce and position sensors on the finger. Building on this work, in [RUCB11], theyaddress non-prehensile planar sliding. While their framework is able to success-13fully predict what happens during sliding and makes some proper adaptations, onemain drawback of their approach is model-based, which means the object infor-mation must be known in advance, such as box dimensions and contact surface.Although their approach is general enough to apply to a large number of objects,methods using an apriori model together with parameter identification suffer fromthe need to manually develop the appropriate mathematical descriptions, which canbe time consuming and difficult [Lyn98].Like [DKK+11], [FMN+03] and [OBA+09], we also take advantage of a learn-ing approach. However, unlike [Lyn99], our focus is on a typical non-prehensilemanipulation task in a model-free situation, which precludes the use of complex apriori dynamical models of the interactions between the object and robot. Allow-ing the robot to learn sensor-to-phenomena mappings through exploration of itsenvironment removes the burden of having to rely on a human expert to develop adetailed apriori model for this mapping. Similarly to [HS04], in which the authorsevaluate time-series of 2D tactile pressure profiles for object recognition, we areusing tactile time series data for estimating aspects of the ongoing task, but withoutany assumptions as to the structure of the manipulable object or the environment.2.4 Dimensionality ReductionDimensionality reduction is the process of reducing the number of random vari-ables under consideration, and can be broadly categorized into feature extractionand feature selection.2.4.1 Feature ExtractionFeature extraction maps the data in high-dimensional space to lower dimensionalspace. New features are computed as linear or non-linear functions of the originaldata. Most feature extraction techniques take an unsupervised [GH+96], [HK10]or self-supervised learning approach [DKS+06], [AMHP07]. The result is a trans-formed and lower-dimensional set of features that more efficiently captures theunderlying structure of the data. Most regression approaches, like principal com-ponent analysis (PCA) and partial least squares (PLS), play such a role.142.4.2 Feature SelectionUnlike feature extraction methods which try to modify the raw data, feature se-lection approaches such as feature similarity [MMP02] and genetic algorithms[HW06] select subsets of the original variables. In cases where the data containsmany redundant or irrelevant features, data analysis such as regression or classi-fication can perform much better and more accurately in a reduced dimensionalspace than that in the original space. We take advantage of this fact and proposeour own feature selection approach which is inspired by the Signal to Noise Ratio(SNR) characterization of a signal.The method proposed in this thesis can be viewed as an application of featureselection, followed by feature extraction, as embodied by the partial least squaresregression model.2.5 Anomaly Detection in Streaming DataSince multidimensional data streams often exhibit considerable structure, high-entropy information is most likely characterized as an anomaly, or outlier, in com-parison to the vast majority of other input data. An anomaly can be defined as anevent or pattern which does not conform to some well-defined notion of normalphenomena. Detecting the existence of anomalies within data streams is an impor-tant topic within both the data mining and machine learning communities [Agg03],[KBDG04], [DKVY06], [YT02], [SWJR07] and has far-reaching applications insuch areas as fault detection, fraud detection, sensor-networks and image process-ing [CBK09].Our framework develops a model of the sensorimotor data over the course ofa prescribed motion and therefore our system can detect anomalies as data pointsthat are in poor correspondence with predicted values. Knowledge of this anomalycan then initiate further action such as autonomous adaptation to cope with thisanomaly or updating the current model to explain this anomaly and therefore betterreflect reality.15Chapter 3The Block Topple-and-Slide TaskThe ?topple-and-slide?manipulation task that we consider in this thesis is shownin Figure 3.1. We evaluate this task in three scenarios: robot with hand, robot withspherical probe and simulation. Within all three scenarios, the end effector pusheson the side of a block, causing it to topple. The rotated block is then pushed backto its starting position against a wall. We consider this task due to its simplic-ity and its potential to reveal environment properties. A prescribed trajectory toachieve this manipulation task is provided by a human expert via kinesthetic teach-ing. Upon replay, the robot tracks the given trajectory in joint space using a stan-dard computed-torque tracking procedure. The given example trajectory should berobust in two ways: repeating the trajectory should cause repeated rotations of theblock, and the same trajectory should remain successful when applied to blockswith variations in mass, friction, and compliance. While in this experimental setupthere is, by definition, no need for any motion adaptation, the ability to estimatetask properties such as mass, friction coefficient, and compliance is, we believe, animportant prelude to being able to achieve online adaptation of motions based onsensorimotor data. The use of a prescribed manipulation trajectory also implies anapproximate correspondence between the current elapsed time within a motion andthe phase of the manipulation. We use the robot-with-hand scenario as an examplein this chapter to define our task.16Figure 3.1: The topple-and-slide task. The first row is the conceptual depic-tion of a clear toppling and sliding action. The second row shows top-pling and sliding for the robot with hand. The third row is this action forthe robot with spherical probe. The last row shows a physically-basedsimulation of the task. 17Each sensory sample is defined bys ? Rns : (?, ??,?, p,w,r,a)where ?, ??,? ? R7 represent respectively the angles, velocities and torque mea-surements of each of the seven joints of the manipulator arm, p ?R7 gives the endeffector?s pose measurements (3D position and 4D quaternion), w ? R6 is the taskwrench measured via the force-torque sensor mounted to the wrist of the robot,r ? R4 are torque measurements for the joints in the robot hand, and a ? R72 aremeasurements coming from arrays of tactile sensors on each of the three fingers,reshaped into a single vector. In our experiments, we also build models for non-prehensile manipulation where the robot hand is replaced by a spherical probe thatis mounted to the force/torque sensor. For this scenario, a reduced version of s isdefined by dropping the measurements associated with the robot hand.A single execution of the manipulation task leads to the capture of a sensori-motor trace, x ?Rn, which consists of the observations of ns sensors each sampledat nt points in time, and then stacked into a single vector as shown in Figure 3.2.Here, n = ns ?nt . All the elements of x are standardized to have a zero mean and avariance of one, as measured across all trials and properties. This ensures that allthe sensorimotor measurements have approximately the same scale, irrespective oftheir origin.Figure 3.2: Visualization for the sensorimotor trace, x. Each uniquely col-ored segment represents one sensory sample s. The whole trace consistsof the observations of ns sensors each sampled at nt points in time.The task properties to be predicted from the sensorimotor trace data are givenby y ? R3 : (m,?,k), where m is the object mass, ? is friction coefficient with18respect to the support surface, and k is the material compliance.In order to learn a predictive model y = f (x), training data is first gathered fornp unique values of the task properties, i.e., sampled variations of mass, compli-ance, and friction. For each setting, the task is repeated nh times. The final datasetis thus defined by the following data pairs:D = {(xp,h,yp,h)}, p ? [1...np],h ? [1...nh].where p is each property variation and h is each trial. A visualization of the dataset D is shown in Figure 3.3.Figure 3.3: Visualization of the sensorimotor data stream D. Each patch witha unique color represents the sensory sample s at one time phase. Thefull horizontal dimension represents one sensorimotor trace x for oneproperty variation with n sensor readings, which is stacked into a singlevector with ns sensors each sampled at nt points in time. The verticaldimension represents all np variations of property set. Lastly, the depthdimension represents all nh trials on one same task property variation.19Chapter 4Prediction ApproachesSupervised learning methods provide a variety of models to make predictions giveninput data that is paired with corresponding output response values that we wish topredict. Two types of methods are considered in our work: probabilistic classifica-tion models and linear regression models.4.1 Probabilistic Classification ModelA probabilistic model is defined by its sample space, events within the samplespace, and probabilities associated with each event. We first consider a NaiveBayes Classifier (NBC), which is a simple probabilistic classifier with conditionalindependence assumption based on Bayes theorem and the maximum posteriorihypothesis. While naive Bayes models may poorly predict calibrated probabilitiesdue to its unrealistic independence assumption [CNM06], it has been reported asa robust probabilistic method in robotic manipulation to handle both spatial andtemporal variabilities of motion [SDKN06], [ITN03], [KTN08].A principal reason for using Naive Bayes Classifier as a baseline method is theassumption that we have no knowledge about which sensors are most informativefor making a desired prediction. Hence, using the likelihood of all the sensoryobservations is a plausible way to judge how likely a particular estimated prop-erty value should be. Given these likelihoods, we can use maximum likelihoodestimation (MLE) to choose the final estimated value. We define our specified20probabilistic model as follows:P(yp|xi) ??ilnP(xi|yp)where P(xi|yp) is given by a normal distribution.Then the model of final estimation is given byy? = argmaxp?[1...np]P(yp|xi)Note that this model restricts the predicted value of y to be one of the previouslyobserved values, i.e., y ? yp.4.2 Linear RegressionThe goal in regression is often to learn a relationship between response outputs yand input explanatory variables X or, more specifically, to predict y? for a givenvalue of X. The linear model either assumes that the regression function is linear,or that the linear model is a reasonable approximation. In our work, because ofthe high dimensionality of raw inputs, we take the second assumption that a linearmodel is a sufficient approximation to reveal the true underlying relationship be-tween the principal features which characterize the sensorimotor data streams andthe properties which define the environment.Under this assumption, we have a vector of inputs X ? Rn and want to predicta vector of outputs y ? Rm, where n is the number of input variables, or features,and m is the number of output parameters, or properties in our case. In general, alinear approximation model which relates X to y has the following form [HTF01]:y =n?j=0x j? j = X ?? ,where ? ? Rn?m is the matrix of model coefficients, or weights, and x0 = 1. Thevariables x j can come from different sources such as known quantitative inputs andinteractions between variables (e.g.: x3 = x1 ?x2). No matter what the source of the21x j, the model is linear in the parameters.Unlike the naive Bayes estimation, linear regression is able to predict arbitrarycontinuously-valued numerical values rather than only selecting among severalknown candidates. We evaluate three linear regression approaches, Least Squares(LS) and Principal Components Regression (PCR), Partial Least Squares (PLS).4.2.1 Least SquaresThe goal of least squares (LS) regression is to seek the linear function of X thatminimizes the sum of squared residuals from y, as shown in Figure 4.1. For moredetails, see [HTF01].Figure 4.1: Linear Least Square fitting with X ? R2 [HTF01].LS regression treats each input as being important and finds the regressionmodel that minimizes the global error. This approach is demonstrated to success-fully model inverse kinematics for a robot manipulator [CSE94] as well as to en-able soft-finger stiffness and contact in robotic grasping and manipulation [KY04].However, LS regression is vulnerable with respect to outlying observations andhigh correlated input variables. Discordant noise mixed in the input data can causethe linear model to deviate far from the true underlying model. LS regression isalso problematic if the number of data points is not significantly larger than the22number of input dimensions. This problematic scenario occurs for some of thedatasets studied in this thesis.4.2.2 Principal Components RegressionPrincipal component regression (PCR) is a method based on principal componentanalysis (PCA) using derived input directions, which is widely used in robotic ma-nipulation such as learning walking gaits [GCR06], object recognition [AAV+08],[CBSF09] and dexterous grasping [KC02]. The purpose of PCR is to estimate thevalues of a response variable in the bases of selected principal components (PCs)of the explanatory input variables [Fil01]. Part or whole of PCs, each of which is alinear combination, Zm of original columns of X, are used in place of the X as inputexplanatory variables. Taking only a subset of PCs also reduces the dimensionalityof regressors while maximizes the variance of the data in new dimensional space,as shown in Figure 4.2. By means of the uncorrelated characteristic of PCs, PCRalso avoids the issue that explanatory input variables are often highly correlatedwhich may cause inaccurate estimates of LS regression coefficients.23Figure 4.2: Principal components of X ? R2 [HTF01]. The largest principalcomponent is along the direction which maximizes the variance of theprojected data.4.2.3 Partial Least SquaresPartial least squares (PLS) regression is a technique that fits linear models usinga hierarchy of univariate regressions along selected projections called latent vari-ables [SAV02]. The latent variables, or scores, are a set of orthogonal factors ex-tracted on the basis of maximizing the covariance between linear combinations ofinput variables X and response y which have the best predictive power. Therefore,24PLS is particularly helpful when facing a (very) large set of explanatory input vari-ables by reducing it to a smaller number of latent variables [Abd03]. In contrast tosimilar regression techniques such as PCR which focus on modeling the directionsof high variance in X, PLS also considers the correlation of response y and inputvariables X for its construction by using y (in addition to X) for its construction.A schematic outline is shown in Figure 4.3 and more mathematical details can befound in [HTF01].Figure 4.3: Schematic outline of PLS regression. The extracted input vari-ables X S (also referred to as predictor scores) are used to predict ex-tracted response scores y S and then the predicted y S are used to con-struct predictions for the response. The coefficient matrix ? can also bederived via this process.? = PLS(y, X).A hierarchical classification of the above is shown in Figure 4.4. Table 4.1 is asummarization of the pros and cons for each method.25Figure 4.4: Prediction approach hierarchyApproach Continuously-valued predictions High variant inputs High correlated inputs and outputsNBC No No YesLS Yes No NoPCR Yes Yes NoPLS Yes Yes YesTable 4.1: Comparisons of different prediction approaches26Chapter 5Structured PredictionMethodologiesGiven a new sensorimotor trace, x, which consists of nt samples of s, we wish toestimate the task properties y. While this could be an obvious application of linearregression, this is in practice problematic because the training data for our scenarioconsists of a relatively small number of data points (low hundreds) embedded in ahigh dimensional space: x can contain thousands of sensorimotor measurements.To cope with these issues, we develop a two-stage solution that consists of (i)selecting the most relevant input features from x, and (ii) using partial least squares(PLS) to further learn a more compact latent linear subspace that is well suited tothe estimation task.5.1 System OverviewAn overview of the system is given in Figure 5.1. It takes four steps to develop amodel:? Data Collection ? After teaching the prescribed trajectory, we replay it onmultiple variations of the mass, friction coefficient, and compliance of themanipulated object so that we collect a series of sensorimotor data streams{(xp,h,yp,h)} across all np property variations and nh repetitions, where xp,his each sensorimotor trace and yp,h is each property variation value. The27specific variables defines x and y are described in chapter 3.? Preprocessing ? Before we can build an appropriate model, however, wemust first whiten the data to avoid any potential bias caused by the varietyof units used across the sensors, which will typically not be comparable.For example, the wrist position data as measured in M is likely to be small,whereas the finger force data as measured in N can be large during sliding.We therefore standardized the sensorimotor data to have a mean of 0 and avariance of 1.? Feature Selection ? The reduced feature vector is extracted by selecting asubset of important X?. These are identified using the Task Variance Ratioas will be defined shortly.? Model Training ? The predictive models are developed by applying partialleast squares model to the reduced feature sets and the know property values.Figure 5.1: System Overview5.2 Feature Selection via the Task Variance RatioTo cope with the high dimensional nature of the sensorimotor trace data, we firstdescribe a selection mechanism that is well suited to identifying specific sensorsand sample times that are likely to be informative for the estimation problem. Theintuition is that we wish to identify features in x that exhibit small variation acrossrepeated trials for the same task properties and exhibit larger variations when mea-sured across all trials and all task property values. Accordingly, we define the Task28Variance Ratio (TVR), ?, as?=Vp,h/Vh,where Vh models the intra-task-property variance of a given element of x, and Vp,hmodels the inter-task-property variance of the same element. Specifically,Vp,h = ?p,h,i.e., the variance of the given element as computed using all available measure-ments (i.e., all properties p and trials h) in D, andVh = ?h,i.e., the mean of the np variances that can be computed when measurements aregrouped according to task properties. We take a large value of ? to be indicative ofa good feature, as it implies that variation occurs as a function of changes of taskproperty values rather than noise observable between repeated trials in the sameenvironment.The feature selection is then implemented using a simple threshold function toproduce a reduced vector x?:X? = {xi|?i > ?min}The resulting reduced dataset is given byD? = {(x?p,h,yp,h)}, p ? [1...np],h ? [1...nh].5.3 Property Estimation with Partial Least SquaresAs introduced in section 4.2.3, partial least squares (PLS) regression is an effectiveestimation method in high-dimensional settings such as the one encountered here.Data from multiple sensors is likely to be highly correlated ? for example, we canassume that the force detected by the force/torque sensor in the wrist is correlatedwith the pressure readings given by the tactile sensors at the fingertips. In addition,there is likely to be significant correlation between sensor readings across time.29Lastly, there are correlations between the input dimensions and output dimensionsthat can readily be exploited. To this end, we find that the partial least squares(PLS) method is superior to other predictive algorithms which do not leverage thiscorrelation in their calculations, e.g., naive Bayes classifier, least squares regres-sion, and principal components regression. In the results chapter, we report onexperiments and comparisons using those approaches in order to demonstrate thatin practice PLS method is the best choice for our task.The PLS algorithm provides us with an estimated weighting matrix ? for eachof our reduced feature dimensions:? = PLS(D?).We use the PLS implementation provided by the Matlab function plsregress. Fi-nally, we use ? at runtime to estimate environment properties:y? = ? ?x?Furthermore, a fixed subset of the largest computed partial least squares coef-ficients ? can be used for the final estimation instead of the full set. We will showthis capability in section Online EstimationOnce the model has been established using the above steps, it can be run in real-time on-board the robot as it collects novel sensorimotor stream data during theblock topple-and-slide task. As one would expect, the prediction of environmen-tal properties is most accurate once the robot has accumulated sensor readingsthroughout the entire motion. However, we are also able to make predictions in anon-demand fashion, without having to wait until the entire motion has completed.We start by parsing the entire motion into a series of key time-phases, t?. We de-fine each t? ? T as a time-phase wherein at least K sensors have received a ? valuelarger than the given threshold value. This ensures that we make predictions onlyonce it is actually plausible to do so. We then build separate prediction modelsfor each new time step t?, where each prediction model takes into account all the30information that is available to date.31Chapter 6Experiments and ResultsIn this chapter, we validate the automatic feature selection and model-free propertyestimation method on the block topple-and-slide task performed by an autonomousrobot which is shown in Figure 3.1. The task involves non-prehensile manipulationof a flexible object, with multiple contact phases occurring between the object, therigid end-effector, and the rigid environment throughout the task.The goal of the manipulation task is to topple the block and bring it to itsfinal resting configuration. In this section we briefly: (1) describe the experimentalsetup; (2) define the block topple-and-slide task; (3) introduce our experimentalprocedure in more detail; and (4) discuss the obtained results.6.1 Robotic Configuration and Experimental TestbedRobot experiments are conducted using a 7 DOF Barrett WAM robot arm withattached 4 DOF Barrett BH-280 Hand, built by Barrett Technology, Inc. A 6-axisforce-torque sensor is mounted to the wrist of the arm. The robot hand is equippedwith tactile arrays on the fingers and palm. Joint torque sensors are embedded ineach of the 3 fingers. Position control of the arm occurs at 500Hz and all sensorsare sampled at 125 Hz.Example trajectories are introduced to the robot via a kinesthetic teach-and-play interface. The system records pose estimates of the arm, i.e., joint angles, atthe rate of 500 Hz and the result is saved for future playback.32Our real-time control framework runs on top of the libbarrett real-time systemslibrary, and is used during demonstration and autonomous execution. We also useit to record and play back data streams that are time-synchronized with the motion.The block used for this experiment is a rectangular prism made of medium-density polyethylene foam, with length 48.5 cm, width 10.5 cm and height 10.5 cm.Two parallel walls of length 28.5 cm and width 6.5 cm are used to prevent theblock from sliding sideways out of the workspace. The distance between the wallsis 49 cm. As shown in Figure 3.1, a wall is used to limit the final sliding motionand leave the block in its original location, ready to be toppled again. The wallsare lined with paper to decrease the coefficient of friction between the block andthe walls for smoother operation. For the robot-with-hand scenario, task propertysets are organized by the Cartesian product of the two sets of property values formass Pm and friction coefficient P? . These values are shown in Table 6.1. Onthe other hand, for the robot with spherical probe scenario, task property sets areorganized by the Cartesian product of three sets of property values for mass Pm,friction coefficient P? and compliance Pc. These values are shown in Table 6.2.Pm (kg)0.425 0.650 0.875 1.101.325 1.550 1.775 2.000P? 0.441 0.505 0.616Table 6.1: Task property for robot with hand.Pm (kg)0.425 0.650 0.875 1.101.325 1.550 1.775 2.000P?0.441 0.505 0.6160.768 0.911 1.136Pc (mm/N) 0.294 2.484 0.978Table 6.2: Task property for robot with spherical probe.We measure the mass of the block by weighing it using a scale. To determinethe friction coefficient, we place the block on an inclined plane and increase theangle ? of inclination of the plane until the block begins to slide. At this critical33moment, the maximum static friction is in effect, which we assume to be approx-imately identical to kinetic friction. The corresponding critical angle shown inFigure 6.1 is related to friction coefficient ? , i.e. ? = tan? . The compliance isdefined asc =?Fwhere, ? is the displacement produced by the force along the same degree of free-dom and F is the force applied on the body. We read the force magnitude fromrelated sensor on the arm and measured the displacement using Vernier calipers.The materials used to change the mass, friction coefficient and compliance in ourrobotic experiment are shown in Figure 6.2.Figure 6.1: Testing friction coefficient using an inclined plane34(a) Mass: 0.225 kg each. The small boxes are filled with sand, placed inside a larger tube.(b) Friction: paper, plastic, wood, fine sandpaper, course sandpaper, towel (left-to-right)(c) Compliance: ethafoam, seafoam, greyfoam (left-to-right)Figure 6.2: Property setup for robots356.2 Simulation Configuration and Experimental TestbedSimulated experiments are performed on a desktop with dual Intel 2.66GHz CPUand 6GB of RAM running Microsoft Windows 7 SP1. Rendering is based on Qt-4.8.3 with additional plugins such as libQGLViewer. The dynamics simulation isimplemented using Open Dynamics Engine (ODE) 0.12 and the simulation sam-pling time step is 0.002 s.The prescribed trajectory of the end effector is trained via a kinesthetic teach-and-play interface interacted with mouse to set the desired position of end effector.Due to the 2D constraint of mouse movement, the prescribed trajectory is actu-ally constrained to 2D space. However, our physically-based simulation is run in3D. We apply Cyclic-Coordinate Descent Inverse Kinematics (CCD IK) [LC98]to calculate desired joint angle and use a simple PD controller to compute jointtorque.The block used in simulation is a rectangular prism with length 25.0 cm, width7.0 cm and height 7.0 cm. As shown in Figure 3.1, a wall of height 5.0 cm is usedto stop the final sliding motion and leave the block in its original location, ready tobe toppled again. The simulated robot has five joints and a ball hand, constrainedby 6 DOF. The fully extended length of the arm is 85.0 cm. Figure 6.3 shows oursimulation environment. In our simulation experiments, environmental propertysets are organized by the Cartesian product of two sets of property values for massPm and friction coefficient P? . These values are shown in Table 6.3.Pm (kg)0.610 0.635 0.660 0.6850.710 0.735 0.760 0.785P? 0.4 0.5 0.6 0.7 0.8Table 6.3: Task property for simulationFigure 6.3: Simulation environment366.3 Procedure6.3.1 Obtain the Task TrajectoryIn the robot experiments, a human expert demonstrates the topple-and-slide task(Figure 3.1) via kinesthetic teaching. The robot base is attached to the table tokeep it fixed. The motion is then manually tuned so that the prescribed trajectorysucceeds for the topple-and-slide task for a variety of combinations of block mass,coefficient of friction, and compliance (see section 6.1). The demonstrating userperforms the task in about 6 s for robot with hand and 3.5 s for robot with sphericalprobe.For the simulation scenario, a user teaches the topple-and-slide task (Figure 3.1)by using the mouse to guide the desired position of the probe. The position of thecursor indicates where the user expects the end effector move to. The prescribedtrajectory is run on a variety of combinations of block mass and surface frictioncoefficient (see section 6.2). The demonstrating user performs the task in about9.2 s.6.3.2 Obtain the Sensorimotor Training Data, DAs described in section 3, the prescribed motion is repeated over a series of trialsh ? [1 ? ? ?nh] for each property set p ? P, yielding the complete raw sensorimotordata set D. The sensory data is resampled to 5 Hz using a box filter. Task parame-ters ns, nt , nh and np for each experiment scenario are shown in Table 6.4.Configuration ns nt nh npRobot with hand 110 15 20 24Robot with spherical probe 28 18 20 144Simulation 16 24 20 40Table 6.4: Task parameters376.3.3 Feature SelectionFollowing the equations in section 5.2, we compute ? for each element in x. Weselect ?min so as to select a fixed percentage of the features. We evaluate variousvalues for this in the next section. We find that in practice good performance isachieved with 25%.6.3.4 Partial Least Squares ModelingUsing PLS, we obtain the ? coefficients. We can make the representation morecompact by further choosing only the ? coefficients of largest magnitude. Wefurther evaluate this in section Online EstimationAs described in section 5.4, there is no point building a prediction model if notenough of the selected informative features have been observed yet. We wait untilat least K selected features have been observed before making any predictions. Inpractice, we choose K = ns/10. By training separate models in this fashion, weare able to make predictions as soon as the robot enters any phase of the motioninvolving selected features, as shown in Figures 6.8 and ResultsIn what follows below, we present the results for topple-and-slide experiments forrobot with hand, robot with spherical probe, as well as the simulation. A summaryof all the relevant figures in this section is shown in Table 6.5. We also encouragethe reader to watch the supplemental video associated with this thesis.Configuration FigureRobot with hand 6.6, 6.8a, 6.9a, 6.10aRobot with spherical probe 6.4, 6.5a, 6.7, 6.8b, 6.9bSimulation 6.5b, 6.8c, 6.9c, 6.10bTable 6.5: Figure illustration386.4.1 Feature Selection PerformanceTVR selection helps focus attention on specific sensors and times that are partic-ularly likely to provide useful information. Figure 6.4 illustrates the selected fea-tures for the topple-and-slide task as executed by the robot arm with the sphericalprobe as an end effector. The intensity of each gray rectangular region is propor-tional to its ? value. The clusters of x? along the time axis can be interpretedas defining important sensorimotor events in the task sequence, which the robotshould pay most attention to. The yellow shaded region corresponds to the top-ple phase and the blue shaded region corresponds to the slide phase. The motionphases where the arm is not in contact with the object are identified as being unim-portant, as are the phases that mark the beginning and end of both of the topple andsliding phases. In terms of sensors, the joint velocities, jvn, are generally identi-fied as being unimportant, with the exception of joint 6. Joints 2, 4, and 6 providetask-relevant information in their sensed torques and positions. Similar results arealso obtained with the full hand attached to the robot arm, in which case there areover a hundred sensors sampled across 18 phases of the motion. With the hand inplace, the key sensors are determined as being the task wrench w, as measured bythe force-torque sensor, the fingertip torques, c, and the fingertip tactile readings, a.39Figure 6.4: Visualization of features selected according to the task varianceratio, ?, of joint position, velocity, torque, and Cartesian pose measure-ments.To determine the impact of the TVR feature selection, we compare mass es-timates obtained using the inclusion of all features, i.e., no feature selection, andthose obtained when TVR feature selection is used to select a subset of the the orig-inal features. This is applied to the manipulation task as executed using robot withspherical probe and simulations. In both cases, a non-reduced partial least squaresmodel is constructed and leave-one-out-cross validation (LOOCV) test is consid-ered for evaluating the performance. As shown in Figure 6.5, the result producedusing the significantly reduced subset of input features is better than that obtainedwhen using all the features.40(a) Robot with hand: 25% TVR feature selection(b) Simulation: 15% TVR feature selectionFigure 6.5: Effect of TVR feature selection on mass estimation416.4.2 Beta Reduction PerformanceIf desired, a fixed subset of the largest computed partial least squares coefficientscan be used for the final estimation, instead of the full set, ? . For example, forrobot with hand scenario in Figure 6.6, we see that a 40% reduction in the numberof coefficients of ? yields only a marginal reduction in the quality of the estimation.Figure 6.6: Effect of partial least squares dimensionality reduction on massestimation.6.4.3 Validation of PLSTo validate our choice of partial least squares (PLS), we compare the results againstthree other methods mentioned in Chapter 4: naive Bayes classifier (NBC), leastsquare regression (LS) and principal component regression (PCR). For NBC, a newsensorimotor trace is treated as input to a classification problem, and the classifieris constructed using a naive Bayes that assumes that all features in x are inde-42pendent. Using the repeated trials for the given set of task properties, a normaldistribution is constructed for each element of x, and the likelihood of a new valueof x belonging to the same class is simply modeled as the product of the individualelement likelihoods. The task properties of the most likely class are then returnedas the estimate. For LS, we regularize the solution using ridge regression. In PCR,only correlations in the input space are considered when establishing a linear spacefor use during regression. All four methods are evaluated using leave-one-out-cross validation (LOOCV) test on the robot with spherical probe scenario, and areapplied to x?, i.e., after TVR feature selection, like shown in Figure 6.7. The re-sults for mass estimation show that PLS yields the best estimates, with respectivemean errors for PLS, PCR, LS and NBC of 0.0333, 0.0564, 0.0849, and 0.2826, asmeasured in kg, respectively.Figure 6.7: LOOCV for mass estimation on NBC, LS, PCR and PLS.436.4.4 Online Property EstimationThe task properties can be estimated online by precomputing multiple PLS models,each of which spans from the start of the motion to a particular time in the motion.Figures 6.8 and 6.9 illustrate the result for the mass and friction coefficient esti-mates of robot with hand, robot with spherical probe and simulation respectively.The estimates improve as the motion progresses and more selected features areobserved.(a) Robot with hand(b) Robot with spherical probe(c) SimulationFigure 6.8: Online estimation of the mass from sensorimotor data. The esti-mates of mass are shown at various blue points throughout the motion.The dotted red horizontal line denotes the actual mass of the block.44(a) Robot with hand(b) Robot with spherical probe(c) SimulationFigure 6.9: Online estimation of the friction coefficient from sensorimotordata. The estimates of friction coefficient are shown at various bluepoints throughout the motion. And the dotted red horizontal line denotesthe actual friction coefficient.6.4.5 Qualitative SimilarityWe also explore the similarity of sensorimotor events that are identified as beingimportant in both the simulation and the real robot. In Figure 6.10, we can obvi-ously find that both of these two setups generate clear toppling and sliding phases,which are qualitatively similar but quantitatively dissimilar. There is also signifi-45cant qualitative similarity in the selected set of sensors that are identified to be themost informative, according to the TVR. The similarity of estimation accuracy isalso revealed through online estimation performance, as shown in Figures 6.8 and6.9.46(a) Robot with hand(b) SimulationFigure 6.10: Qualitative similarity in task variance ratio, ?476.5 DiscussionThe framework demonstrates the ability to use uninterpreted sensory data streamscollected during a manipulation task in order to make reliable predictions aboutquantities that cannot be visually observed, i.e., mass, friction coefficient, andcompliance. A caveat is that this assumes the existence of the relevant trainingexamples. Sensors or motion phases that are observed to be noisy are readily dis-counted by the method. The results show that the task variance ratio, ?, provides asimple means for feature selection, and which furthermore significantly improvesthe resulting partial least squares estimates. Important motion phases and sensorscan readily be identified. A limitation is that ? can also be misleading, in the caseof noise-free features that also exhibit significant non-linearities with respect to theproperties being estimated. Partial least squares outperforms the other benchmarkestimation methods for the task, and also provides a further opportunity for dimen-sionality reduction if desired. Taken in combination with the feature selection, theresulting framework is simple and effective. The qualitatively similarity betweenreal robot and simulation is evident in our model-free approach, whereas the quan-titative dissimilarity is due to the absence of an accurate model of robot kinematicsand dynamics.The same framework can also be leveraged in order to detect anomalous events.Rapid changes in the estimated task properties, such as a change in the mass, is asignal of an anomaly. Implicit in the computation of ?h and ?p,h is a model ofwhat value a sensory feature should have at a given point in the motion. Thisallows a motion anomaly to be signaled if a number of sensors each begin to signalanomalous values at a given point in time, or a sensor anomaly to be signaled if asingle sensor begins to consistently produce anomalous readings.48Chapter 7Conclusions and Future WorkWe have introduced a model-free approach to the estimation of task properties suchas mass, friction, and compliance during the course of a non-prehensile manipu-lation task. Given appropriate data for example manipulations with known taskproperties, the method can work to extract the needed information from uninter-preted sensorimotor data that is collected over the course of a new manipulation.We demonstrate that important features (sensors and sample times) can be identi-fied and allow for better estimates to be made using as little as 10% of the originaldata, when used in concert with a partial least squares (PLS) estimator. Further re-ductions of dimensionality can then be achieved using the linear latent space modelimplied by the PLS.One significant limitation of the predictive framework is the need for a pre-scribed motion that can already succeed at the task despite variations in the taskproperties that we seek to estimate, such as the mass, friction, and compliance. Animportant direction for future work will be to investigate the tight integration of es-timation and adaptation in the framework. With knowledge (learned or provided)of how to adapt the topple-and-slide task for heavier or more compliant blocks,this could readily be used to enlarge the range of variations that can be coped with.A related limitation is that our current sensorimotor features are all time-indexed,i.e., there is an assumption that the current phase of the motion is tightly coupledto the current time. In future work, we would like to couple the phase estimatemore tightly to the actual motion via the available sensorimotor observations. A49last limitation is that because of the model-free nature of the current approach, theestimation procedures do not generalize well to changes in the task kinematics ordynamics. We aim to develop parameterized versions of the estimation model inorder to allow for such generalization.As another avenue of future work, we are interested in building a more accuratemodel of robot kinematics and dynamics used in simulation in order to achievequantitatively similarity. The developed models could be tested on a system onwhich ground-truth dynamics information were available. An adaptation strategycould ideally be directly transferred from a simulation onto a real robot with littlemodification. It also gives the system a means of training in the absence of physicalmovement. This would give the robot a capacity similar to mental rehearsal asevident in humans [SR64].50Bibliography[AAV+08] Tamim Asfour, Pedram Azad, Niko Vahrenkamp, KristianRegenstein, Alexander Bierbaum, Kai Welke, Joachim Schroeder,and Ruediger Dillmann. Toward humanoid manipulation inhuman-centred environments. Robotics and Autonomous Systems,56(1):54?65, 2008. ? pages 23[Abd03] Herve? Abdi. Partial least square regression (pls regression), 2003.? pages 25[ADA89] Eric W Aboaf, Steven M Drucker, and Christopher G Atkeson.Task-level robot learning: Juggling a tennis ball more accurately. InRobotics and Automation, 1989. Proceedings., 1989 IEEEInternational Conference on, pages 1290?1295. IEEE, 1989. ?pages 12[Agg03] Charu C Aggarwal. A framework for diagnosing changes inevolving data streams. In Proceedings of the 2003 ACM SIGMODinternational conference on Management of data, pages 575?586.ACM, 2003. ? pages 15[AH93] Sami Atiya and Gregory D Hager. Real-time vision-based robotlocalization. Robotics and Automation, IEEE Transactions on,9(6):785?800, 1993. ? pages 10[AMHP07] Anelia Angelova, Larry Matthies, Daniel M Helmick, and PietroPerona. Dimensionality reduction using automatic supervision forvision-based terrain learning. In Robotics: Science and Systems.Citeseer, 2007. ? pages 14[CBK09] Varun Chandola, Arindam Banerjee, and Vipin Kumar. Anomalydetection: A survey. ACM Comput. Surv., 41(3):15:1?15:58, July2009. ? pages 1551[CBSF09] Alvaro Collet, Dmitry Berenson, Siddhartha S Srinivasa, and DaveFerguson. Object recognition and full pose registration from asingle image for robotic manipulation. In Robotics and Automation,2009. ICRA?09. IEEE International Conference on, pages 48?55.IEEE, 2009. ? pages 23[CNM06] Rich Caruana and Alexandru Niculescu-Mizil. An empiricalcomparison of supervised learning algorithms. In Proceedings ofthe 23rd international conference on Machine learning, pages161?168. ACM, 2006. ? pages 20[CSE94] Stefano Chiaverini, Bruno Siciliano, and Olav Egeland. Review ofthe damped least-squares inverse kinematics with experiments on anindustrial robot manipulator. Control Systems Technology, IEEETransactions on, 2(2):123?134, 1994. ? pages 22[Dia10] Rosen Diankov. Automated Construction of Robotic ManipulationPrograms. PhD thesis, Carnegie Mellon University, RoboticsInstitute, August 2010. ? pages viii, 2[DKK+11] Renaud Detry, Dirk Kraft, Oliver Kroemer, Leon Bodenhagen, JanPeters, Norbert Kru?ger, and Justus Piater. Learning grasp affordancedensities. Paladyn, 2(1):1?17, 2011. ? pages 13, 14[DKS+06] Hendrik Dahlkamp, Adrian Kaehler, David Stavens, SebastianThrun, and Gary R Bradski. Self-supervised monocular roaddetection in desert terrain. In Robotics: science and systems,volume 38, 2006. ? pages 14[DKVY06] Tamraparni Dasu, Shankar Krishnan, Suresh Venkatasubramanian,and Ke Yi. An information-theoretic approach to detecting changesin multi-dimensional data streams. In In Proc. Symp. on theInterface of Statistics, Computing Science, and Applications.Citeseer, 2006. ? pages 15[DMVS10] Ravinder S Dahiya, Giorgio Metta, Maurizio Valle, and GiulioSandini. Tactile sensingfrom humans to humanoids. Robotics, IEEETransactions on, 26(1):1?20, 2010. ? pages 11[Ell87] Randy Ellis. Acquiring tactile data for the recognition of planarobjects. In Robotics and Automation. Proceedings. 1987 IEEEInternational Conference on, volume 4, pages 1799?1805. IEEE,1987. ? pages 1152[FHD98] Holger Friedrich, J Holle, and Ru?diger Dillmann. Interactivegeneration of flexible robot programs. In Robotics and Automation,1998. Proceedings. 1998 IEEE International Conference on,volume 1, pages 538?543. IEEE, 1998. ? pages 12[Fil01] P Filzmoser. Robust principal component regression. Computerdata analysis and modeling. Robust and computer intensivemethods. Belarusian State University, Minsk, pages 132?137, 2001.? pages 23[FMN+03] Paul Fitzpatrick, Giorgio Metta, Lorenzo Natale, Sajit Rao, andGiulio Sandini. Learning about objects through action-initial stepstowards artificial cognition. In Robotics and Automation, 2003.Proceedings. ICRA?03. IEEE International Conference on,volume 3, pages 3140?3145. IEEE, 2003. ? pages 13, 14[GCR06] David B Grimes, Rawichote Chalodhorn, and Rajesh PN Rao.Dynamic imitation in a humanoid robot through nonparametricprobabilistic inference. In Robotics: science and systems, pages199?206. Cambridge, MA, 2006. ? pages 23[GH+96] Zoubin Ghahramani, Geoffrey E Hinton, et al. The em algorithm formixtures of factor analyzers. Technical report, Technical ReportCRG-TR-96-1, University of Toronto, 1996. ? pages 14[GLP84] W Eric L Grimson and Tomas Lozano-Perez. Model-basedrecognition and localization from sparse range or tactile data. TheInternational Journal of Robotics Research, 3(3):3?35, 1984. ?pages 11[HC93] Robert D Howe and Mark R Cutkosky. Dynamic tactile sensing:Perception of fine surface features with stress rate sensing. Roboticsand Automation, IEEE Transactions on, 9(2):140?151, 1993. ?pages 11[HGCB08] Micha Hersch, Florent Guenter, Sylvain Calinon, and Aude Billard.Dynamical system modulation for robot learning via kinestheticdemonstrations. Robotics, IEEE Transactions on, 24(6):1463?1467,2008. ? pages 12[HK10] Albert Hein and Thomas Kirste. Unsupervised detection of motionprimitives in very high dimensional sensor data. In Proceedings of53the 5th Workshop on Behaviour Monitoring and Interpretation,BMI, 2010. ? pages 14[How93] Robert D Howe. Tactile sensing and control of roboticmanipulation. Advanced Robotics, 8(3):245?261, 1993. ? pages 11[HS04] Gunther Heidemann and Matthias Schopfer. Dynamic tactilesensing for object identification. In Robotics and Automation, 2004.Proceedings. ICRA?04. 2004 IEEE International Conference on,volume 1, pages 813?818. IEEE, 2004. ? pages 14[HTF01] Trevor. Hastie, Robert. Tibshirani, and J Jerome H Friedman. Theelements of statistical learning, volume 1. Springer New York,2001. ? pages iii, ix, 21, 22, 24, 25[HW06] Cheng-Lung Huang and Chieh-Jen Wang. A ga-based featureselection and parameters optimizationfor support vector machines.Expert Systems with applications, 31(2):231?240, 2006. ? pages15[ITN03] Tetsunari Inamura, Iwaki Toshima, and Yoshihiko Nakamura.Acquiring motion elements for bidirectional computation of motionrecognition and generation. In Experimental robotics viii, pages372?381. Springer, 2003. ? pages 20[KBDG04] Daniel Kifer, Shai Ben-David, and Johannes Gehrke. Detectingchange in data streams. In Proceedings of the Thirtieth internationalconference on Very large data bases-Volume 30, pages 180?191.VLDB Endowment, 2004. ? pages 15[KC02] Danica Kragic and Henrik I Christensen. Model based techniquesfor robotic servoing and grasping. In Intelligent Robots andSystems, 2002. IEEE/RSJ International Conference on, volume 1,pages 299?304. IEEE, 2002. ? pages 23[KETJ07] Charles C Kemp, Aaron Edsinger, and Eduardo Torres-Jara.Challenges for robot manipulation in human environments [grandchallenges of robotics]. Robotics & Automation Magazine, IEEE,14(1):20?29, 2007. ? pages iii, 1, 2, 3, 11[KFE96] Ishay Kamon, Tamar Flash, and Shimon Edelman. Learning tograsp using visual information. In Robotics and Automation, 1996.Proceedings., 1996 IEEE International Conference on, volume 3,pages 2470?2476. IEEE, 1996. ? pages 1054[KH97] Reinhard Koeppe and Gerd Hirzinger. Learning compliant motionsby task-demonstration in virtual environments. In ExperimentalRobotics IV, pages 297?307. Springer, 1997. ? pages 12[KI97] Sing Bing Kang and Katsushi Ikeuchi. Toward automatic robotinstruction from perception-mapping human grasps to manipulatorgrasps. Robotics and Automation, IEEE Transactions on,13(1):81?95, 1997. ? pages 12[KPB08] Dov Katz, Yuri Pyuro, and Oliver Brock. Learning to manipulatearticulated objects in unstructured environments using a groundedrelational representation. In In Robotics: Science and Systems.Citeseer, 2008. ? pages 12[KSN10] Ellen Klingbeil, Ashutosh Saxena, and Andrew Y Ng. Learning toopen new doors. In Intelligent Robots and Systems (IROS), 2010IEEE/RSJ International Conference on, pages 2751?2757. IEEE,2010. ? pages viii, 5, 11[KTN08] Dana Kulic?, Wataru Takano, and Yoshihiko Nakamura. Incrementallearning, clustering and hierarchy formation of whole body motionpatterns using adaptive hidden markov chains. The InternationalJournal of Robotics Research, 27(7):761?784, 2008. ? pages 20[KY04] Imin Kao and Fuqian Yang. Stiffness and contact mechanics for softfingers in grasping and manipulation. Robotics and Automation,IEEE Transactions on, 20(1):132?135, 2004. ? pages 22[LC98] Jeff Lander and GRAPHIC CONTENT. Making kine more flexible.Game Developer Magazine, 1:15?22, 1998. ? pages 36[Lee00] Mark H Lee. Tactile sensing: new directions, new challenges. TheInternational Journal of Robotics Research, 19(7):636?643, 2000.? pages 11[Low99] David G Lowe. Object recognition from local scale-invariantfeatures. In Computer vision, 1999. The proceedings of the seventhIEEE international conference on, volume 2, pages 1150?1157.Ieee, 1999. ? pages 10[Low04] David G Lowe. Distinctive image features from scale-invariantkeypoints. International journal of computer vision, 60(2):91?110,2004. ? pages 1055[Lyn98] Kevin M Lynch. Issues in nonprehensile manipulation. In Roboticsthe Algorithmic perspective: The Third Workshop on theAlgorithmic Foundations of Robotics, pages 237?250. Citeseer,1998. ? pages 13, 14[Lyn99] Kevin M Lynch. Toppling manipulation. In Robotics andAutomation, 1999. Proceedings. 1999 IEEE InternationalConference on, volume 4, pages 2551?2557. IEEE, 1999. ? pages13, 14[Mas86] Matthew T Mason. Mechanics and planning of manipulator pushingoperations. The International Journal of Robotics Research,5(3):53?71, 1986. ? pages 13[Mel00] Claudio Melchiorri. Slip detection and control using tactile andforce sensors. Mechatronics, IEEE/ASME Transactions on,5(3):235?243, 2000. ? pages 12[Mel01] Claudio Melchiorri. Tactile sensing for robotic manipulation. InRamsete, pages 75?102. Springer, 2001. ? pages 12[MMP02] Pabitra Mitra, CA Murthy, and Sankar K. Pal. Unsupervised featureselection using feature similarity. IEEE transactions on patternanalysis and machine intelligence, 24(3):301?312, 2002. ? pages15[Mor07] John Mortimer. Bmw creates production triangle in the uk tomanufacture the new mini. Industrial Robot: An InternationalJournal, 34(1):26?31, 2007. ? pages viii, 3[Neu13] Gerhard Neumann. Introduction by the organizers. In NovelMethods for Learning and Optimization of Control Policies andTrajectories for Robotics at ICRA?13,, 2013. ? pages 8[OBA+09] Damir Omrcen, Christian Boge, Tamim Asfour, Ales Ude, andRu?diger Dillmann. Autonomous acquisition of pushing actions tosupport object grasping with a humanoid robot. In HumanoidRobots, 2009. Humanoids 2009. 9th IEEE-RAS InternationalConference on, pages 277?283. IEEE, 2009. ? pages 13, 14[PBS07] C Pepper, S Balakirsky, and C Scrapper. Robot simulation physicsvalidation. In Proceedings of the 2007 Workshop on PerformanceMetrics for Intelligent Systems, pages 97?104. ACM, 2007. ?pages 256[PRKS11] Peter Pastor, Ludovic Righetti, Mrinal Kalakrishnan, and StefanSchaal. Online movement adaptation based on previous sensorexperiences. In Intelligent Robots and Systems (IROS), 2011IEEE/RSJ International Conference on, pages 365?371. IEEE,2011. ? pages 12[RUCB10] Federico Ruiz-Ugalde, Gordon Cheng, and Michael Beetz.Prediction of action outcomes using an object model. In IntelligentRobots and Systems (IROS), 2010 IEEE/RSJ InternationalConference on, pages 1708?1713. IEEE, 2010. ? pages 13[RUCB11] Federico Ruiz-Ugalde, Gordon Cheng, and Michael Beetz. Fastadaptation for effect-aware pushing. In Humanoid Robots(Humanoids), 2011 11th IEEE-RAS International Conference on,pages 614?621. IEEE, 2011. ? pages 13[SAV02] Stefan Schaal, Christopher G Atkeson, and Sethu Vijayakumar.Scalable techniques from nonparametric statistics for real time robotlearning. Applied Intelligence, 17(1):49?60, 2002. ? pages 24[SC98] James C Spall and John A Cristion. Model-free control of nonlinearstochastic systems with discrete-time measurements. AutomaticControl, IEEE Transactions on, 43(9):1198?1210, 1998. ? pages 8[SDK+06] Ashutosh Saxena, Justin Driemeyer, Justin Kearns, Chioma Osondu,and Andrew Y Ng. Learning to grasp novel objects using vision. In10th International Symposium of Experimental Robotics (ISER).Citeseer, 2006. ? pages 10[SDKN06] Ashutosh Saxena, Justin Driemeyer, Justin Kearns, and Andrew Ng.Robotic grasping of novel objects. In Advances in neuralinformation processing systems, pages 1209?1216, 2006. ? pages20[SHC92] Rajeev Sharma, J-Y Herve, and Peter Cucka. Dynamic robotmanipulation using visual tracking. In Robotics and Automation,1992. Proceedings., 1992 IEEE International Conference on, pages1844?1849. IEEE, 1992. ? pages 11[SLL01] Stephen Se, David Lowe, and Jim Little. Vision-based mobile robotlocalization and mapping using scale-invariant features. In Roboticsand Automation, 2001. Proceedings 2001 ICRA. IEEE International57Conference on, volume 2, pages 2051?2058. IEEE, 2001. ? pages10[SLL05] Stephen Se, David G Lowe, and James J Little. Vision-based globallocalization and mapping for mobile robots. Robotics, IEEETransactions on, 21(3):364?375, 2005. ? pages 10[SMH94] Jae S Son, Eduardo A Monteverde, and Robert D Howe. A tactilesensor for localizing transient events in manipulation. In Roboticsand Automation, 1994. Proceedings., 1994 IEEE InternationalConference on, pages 471?476. IEEE, 1994. ? pages 11[SPK02] William D Smart and L Pack Kaelbling. Effective reinforcementlearning for mobile robots. In Robotics and Automation, 2002.Proceedings. ICRA?02. IEEE International Conference on,volume 4, pages 3404?3410. IEEE, 2002. ? pages 6[SR64] Kenneth B Start and Alan Richardson. Imagery and mental practice.British Journal of Educational Psychology, 34(3):280?284, 1964.? pages 50[SWJR07] Xiuyao Song, Mingxi Wu, Christopher Jermaine, and Sanjay Ranka.Statistical change detection for multi-dimensional data. InProceedings of the 13th ACM SIGKDD international conference onKnowledge discovery and data mining, pages 667?676. ACM, 2007.? pages 15[SYKM06] N Sian, K Yoki, Y Kawai, and K Muruyama. Operating humanoidrobots in human environments. In Proceedings of the Robotics,Science & Systems Workshop on Manipulation for HumanEnvironments, Philedelphia, Pennsylvania. Citeseer, 2006. ? pages4[Tat13] Alan Tatourian. Human brain and neuroscience, 2013. ? pagesviii, 7[TLG73] MM Taylor, SJ Lederman, and RH Gibson. Tactual perception oftexture. Handbook of perception., 3:251?272, 1973. ? pages 11[TZPvdP14] Daniel M. Troniak, Chuan Zhu, Dinesh K. Pai, and Michiel van dePanne. Structured predictions from sensorimotor data duringnon-prehensile manipulation, 2014. ? pages iii58[WJ84] G Westling and RS Johansson. Factors influencing the force controlduring precision grip. Experimental Brain Research,53(2):277?284, 1984. ? pages 11[YT02] Kenji Yamanishi and Jun-ichi Takeuchi. A unifying framework fordetecting outliers and change points from non-stationary time seriesdata. In Proceedings of the eighth ACM SIGKDD internationalconference on Knowledge discovery and data mining, pages676?681. ACM, 2002. ? pages 15[ZL12] Yuan-Fei Zhang and Hong Liu. Tactile sensor based varying contactpoint manipulation strategy for dexterous robot hand manipulatingunknown objects. In Intelligent Robots and Systems (IROS), 2012IEEE/RSJ International Conference on, pages 4756?4761. IEEE,2012. ? pages 1359Appendix ASupporting MaterialsA corresponded video is attached to this paper.60


Citation Scheme:


Citations by CSL (citeproc-js)

Usage Statistics



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"
                            async >
IIIF logo Our image viewer uses the IIIF 2.0 standard. To load this item in other compatible viewers, use this url:


Related Items