Probability-Weighted Temporal Registration for Improving Robot Motion Planning and Control Learned from Demonstrations

Abstract

Many existing methods that learn robot motion planning task models or control policies from demonstrations require that the demonstrations be temporally aligned. Temporal registration involves an assignment of individual observations from a demonstration to the ordered steps in some reference model, which facilitates learning features of the motion over time. We introduce probability-weighted temporal registration (PTR), a general form of temporal registration that includes two useful features for motion planning and control policy learning. First, PTR explicitly captures uncertainty in the temporal registration. Second PTR avoids degenerate registrations in which too few observations are aligned to a time step. Our approach is based on the forward-backward algorithm. We show how to apply PTR to two task model learning methods from prior work, one which learns a control policy and another which learns costs for a sampling-based motion planner. We show that incorporating PTR yields higher-quality learned task models that enable faster task executions and higher task success rates.