Recent advances in robot platforms have outpaced our ability to effectively program robots to accomplish useful tasks, often in complex environments that they share with humans. In order for flexible, general purpose robots to become widespread e.g., in teaching skills to children, assisting the elderly, there must be a way of interacting with them beyond programming. Teaching by demonstration offers a potentially powerful and practical approach to realizing the promise of large scale personal robotics in a wide range of applications. In teaching by demonstration, the expert (human), demonstrates the task on different hardware than what the apprentice or student (robot) uses.
The project aims to develop visual feature-based methods that allow robots to teach humans and learn from them by unifying apprenticeship learning, learning by demonstration (or by imitation) and teaching humans, taking into account the differences between experts and apprentices. The resulting system will be evaluated on a PR2 robot (mostly on grasping and manipulation tasks). The scientific advances resulting from the project in learning from demonstration and imitation learning, both general techniques with broad applicability, will greatly simplify the programming of robots which would make it easier for non-expert users to perform this important task which currently requires considerable expertise in robotics as well as computer science.
Broader impacts of the research include development of new robotics curricula, enhanced opportunities for research-based interdisciplinary training at the intersection of computer vision, machine learning, and robotics, outreach activities (including participation in a public school robotics instruction program). All of the results of the research, including publications, open-source software and datasets, will be made freely available to the larger scientific and academic community.
