Combining motion planning and optimization for flexible robot manipulation

摘要

Robots that operate in natural human environments must be capable of handling uncertain dynamics and underspecified goals. Current solutions for robot motion planning are split between graph-search methods, such as RRT and PRM which offer solutions to high-dimensional problems, and Reinforcement Learning methods, which relieve the need to specify explicit goals and action dynamics. This paper addresses the gap between these methods by presenting a task-space probabilistic planner which solves general manipulation tasks posed as optimization criteria. Our approach is validated in simulation and on a 7-DOF robot arm that executes several tabletop manipulation tasks. First, this paper formalizes the problem of planning in underspecified domains. It then describes the algorithms necessary for applying this approach to planar manipulation tasks. Finally it validates the algorithms on a series of sample tasks that have distinct objectives, multiple objects with different shapes/dynamics, and even obstacles that interfere with object motion.