Open Collections

Abstract

Robots have the potential to support humans in diverse domains such as industrial automation, healthcare, and household tasks. However, enabling robots to learn new tasks remains challenging, especially for non-expert users. Learning from Demonstration (LfD) offers a promising solution by allowing robots to learn directly from human demonstrations rather than requiring explicit programming. Despite its potential, the success of LfD hinges on the quality, quantity, and diversity of the demonstrations provided. This thesis addresses the challenges of ensuring high-quality and informative demonstrations to improve learning and generalization. First, it introduces a performance-based framework to evaluate demonstration quality, showing that robot task success and generalization can serve as objective measures. It also proposes a set of consistency-based metrics that can be applied prior to training, enabling the identification and filtering of low-quality demonstrations in advance, thus improving learning outcomes. To further enhance demonstration quality, this thesis explores train-the-trainer methodologies that help novice users without a robotics or programming background provide expert-level demonstrations. Three different training strategies are investigated: discovery training, where users learn through practice and self-reflection; observational training, where users watch expert demonstrations; and kinesthetic training, where users experience expert movements through physical interaction with the robot. User studies validate the effectiveness of these approaches, showing improvements in demonstration quality, and the ability to transfer skills to new tasks. Finally, this thesis explores the informativeness of demonstrations, addressing the issue of data sparsity in LfD. Instead of treating all demonstrations equally, we propose using information entropy to identify the most uncertain regions in the task space, prompting users to provide additional demonstrations where they are most needed. This approach improves learning efficiency by reducing redundant demonstrations while ensuring sufficient data coverage for generalization. Across multiple experiments and user studies, the findings demonstrate that combining quality assessment, targeted user training, and data-efficient selection strategies significantly enhances robot learning performance. This work advances LfD by making it more adaptive, scalable, and accessible to non-experts. Ultimately, it contributes to the development of intuitive human-robot interaction systems where everyday users can teach robots effectively, opening new possibilities for deploying robots in real-world, user-centered environments.