Exploiting Discrete Structure for Learning On-Line in Distributed Robot Systems
Technical Report,05 Jul 2005,31 Mar 2009
University of Massachusetts - Amherst Amherst United States
Pagination or Media Count:
Over the course of this award, we have introduced techniques for adaptive control that exploit discrete structure in the organization of distributed and embedded systems. This general approach was evaluated using a variety of physical platforms including a distributed sensor array, a team of multiple mobile platforms, a bimanual humanoid and a new concept in mobile manipulators. The findings suggest that this approach has some advantages with respect to learning performance, fault tolerance, generalization, transfer, and programming by demonstration. In particular, our cumulative publications over the award period contribute 1 a representation for creating a comprehensive array of closed-loop control circuits, 2 a discrete-event framework for characterizing the state of the dynamical system, 3 and model checking scheme for guaranteeing performance and safety while learning that significantly improves the performance of machine learning techniques in robotics, 4 an intrinsically motivated reinforcement learning technique based on the discrete-event representation, 5mechanisms abstraction in discrete event systems that enhances transfer and provides a means of learning from demonstrations and imitation.