Motion Planning for Highly Constrained Spaces

We introduce a sampling-based motion planning method that automatically adapts to the difficulties caused by thin regions in the free space (not necessarily narrow corridors). These problems arise frequently in settings such as closedchain manipulators, humanoid motion planning, and generally any time bodies are in contact or maintain close proximity with each other. Our method combines the aggressive exploration properties of RRTs, the adaptive sampling domain of DDRRTs, and the intrinsic dimensionality-reduction properties of kd-trees to focus the sampling and searching in the appropriate subspaces. We handle closed-chains and other kinds of constraints in a general way that avoids inverse kinematics computations, if desired. We have implemented the method and show its computational advantages on a variety of challenging examples.