Faster Algorithms for Growing Collision-Free Regions of Constrained Bimanual Configuration Spaces
Abstract
Decomposition-based motion planners (DBMPs) have shown incredible effectiveness, combining the best aspects of sampling-based planners and trajectory optimization. DBMPs first decompose the free space into collections of convex sets and then optimize trajectories contained in the union of these sets. So far, DBMPs have been unable to deal with the curved constraint manifolds that appear in constrained bimanual manipulation tasks natively. Recent work has leveraged a parametrization of the constraint manifold, induced by analytic inverse kinematics, to build collision-free convex sets on such manifolds. We combine this strategy with recent advances in collision-free convex set generation to build regions up to 500x faster. The resulting sets also have simpler descriptions, (usually) larger volume, and a lower fraction in collision.