Robots will be more useful to society as they respond to more natural teaching paradigms similar to the way people learn. It is hard to tell robots what to do, and even harder to tell robots how to handle all possible errors, unanticipated events, and accidents that might happen. This project focuses on enabling robots to understand instructions from users, to fix errors and repair broken processes, objects, and tools, and to improvise and find new and better ways to do tasks. A key part of the research is to enable robots to learn from educational kits and instructional material typically given to children. The researchers plan to use these instructional materials to provide robots with "physical common sense" about how the world works, in a similar way we expect children to learn. The research will make programming robots easier and cheaper, and make robots more useful, particularly for domestic and care robots supporting everyday life activities; repair, construction, and decommissioning robots; and exploration and worker robots in the oceans and space.
Technically, this integrative project addresses a longer-term vision and intellectual challenge of developing a library approach to robot behavior generation and learning. The researchers will build a large-scale long-term intelligent physical system that can repair and improvise desired processes and devices. The system will learn from instruction in the form of a curriculum of training tasks, watching humans do tasks, coaching, practice, less directed play, and reflection. Evaluation of the proposed work will focus on how well robots can perform suggested activities from the instructional material, repair broken processes and devices, and create processes and devices that achieve new task specifications. The following hypotheses will be tested: 1) appropriate libraries can be built in practice and grown over time, 2) relevant experience can be accessed from a large library and combined to exhibit rich behavior comparable to humans, and 3) such a library can support life-long learning for many tasks, rather than a single demonstration of one task in one domain. Key ideas include the use of task and strategy-specific quantitative models, the development of strategy graphs that combine both symbolic descriptions of how to do a task as well as task-level models, and the use of learned simulators to support mental practice, exploration, and learning.
This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.
