Planning the sequencing of movement primitives

(*Work done while at EPFL-VRlab) Neuroscience evidence supports the idea that biological adaptive behavior may utilize combination and sequences of movement primitives, allowing the motor system to reduce the dimensionality of movement control. We present a framework, using sampling-based motion planning, that is able to automatically determine the sequencing of parametric movement primitives needed to execute a given motion task. Our approach builds a search tree in which nodes are configurations reachable with one or more movement primitives, and edges represent valid paths connecting parent and child nodes. The paths are determined by a motion planner that operates in the parameter space of a single movement primitive. The search tree is expanded with A*-like best-first search using greedy problem-specific heuristics. The benefits of our approach are twofold: 1) planning complex motions becomes more efficient in the reduced dimensionality of each movement primitive, and 2) the ability to plan entire motions containing heterogeneous types of constraints, such as collision-free, balanced, alternating support contacts, etc. We present a general framework and several simulation results of statically stable biped walking motions among obstacles. The presented planning capabilities enable robots to better handle unpredicted situations, and can be used as a method of self-organization of higher-level primitives. 1