Incrementally Increasing the Uncertainty-Tolerance of Robotic Manipulation Plans

摘要

Robotic manipulators control is subject to control and sensing errors which isparticularly troublesome in fine motion planning. Several approaches have been taken including propagation of numeric or symbolic uncertainties and generation of plans which use compliant motion to achieve goals in spite of uncertainties. Since approaches have generally sought to guarantee success, plan generation costs are high. We present a new approach to incrementally acquiring uncertainty tolerance in robotic manipulation plans through experience. During plan construction and execution, no reasoning about uncertainty takes place; consequently, plan generation and execution is very fast. However, in response to failures, plans are refined to increase their uncertainty tolerance so as to reduce the future possibility of the encountered failure. The incremental refinement approach has several advantages over guaranteed plans: (1) resulting plans are general and have explicit applicability conditions; (2) plans achieve a savings because they do not explicitly consider uncertainties; (3) savings are obtained over the guaranteed case since often only a subset of all uncertainties lead to failures in practice; and (4) unguaranteed but practical plans can be generated by the incremental approach when they lie outside the scope of the guaranteed planner. To demonstrate our approach we describe an implemented system called GRASPER which learns to grasp novel objects given only imprecise television camera input. No prior model of objects is assumed, nor are the objects required to satisfy a priori constraints on their shapes.