The proposed Phase I development demonstrates a formal control basis for haptically-guided grasping and manipulation. A sensor-based scheme is advanced in which closed-loop control and incremental modeling agents comprise an adaptive sensorimotor mechanism. To react to context dependencies, we will employ reinforcement learning techniques to-approximate optimal control composition strategies. This proposal forms a partnership between ACSIOM Laboratories (AL), Amherst, MA and the Laboratory for Perceptual Robotics (LPR) at the University of Massachusetts in order to transfer technology for autonomous grasping and manipulation developed within the LPR into Navy applications. In Phase I, we will demonstrate a working testbed capable of a range of haptically-guided grasping and manipulation behavior, including mechanisms for accomodating unknown object geometry, for adaptive control compensation and for object recognition. Phase I development will take place on functioning hand/arm testbeds within the LPR. In addition, subsequent Phase II development extends existing techniques for behavioral composition to support automatic acquisition of end-to-end task level problem solving and delivers an integrated architecture for haptically-guided grasping and manipulation to the Navy. This approach structures the behavioral synthesis problem, yields predictable/safe run-time performance, and preserves flexibility while introducing minimal combinatoric complexity in the control composition problem.
Benefits: adaptive control and reinforcement learning for dextrous manipulators low-bandwidth telemanipulation semi-autonomous behabior for hazardous environments and maintenance
Keywords: control learning dexterity haptic telemanipulation control learning dexterity haptic telemanipulation
