A Bayesian Active Learning Approach to Adaptive Motion Planning

Abstract

An important requirement for a robot to operate reliably in the real world is a robust motion planning module. Current planning systems do not have consistent performance across all situations a robot encounters. We are interested in planning algorithms that adapt during a planning cycle by actively inferring the structure of the valid configuration space, and focusing on potentially good solutions. Consider the problem of evaluating edges on a graph to discover a good path. Edges are not alike in value—some are important, others are informative. Important edges have a lot of good paths flowing through them. Informative edges, on being evaluated, affect the likelihood of other neighboring edges being valid. Evaluating edges is expensive, both for robots with complex geometries like robot arms, and for robots with limited onboard computation like UAVs. Until now, we have addressed this challenge via laziness, deferring edge evaluation until absolutely necessary, with the hope that edges turn out to be valid. Our key insight is that we can do more than passive laziness—we can actively probe for information. We draw a novel connection between motion planning and Bayesian active learning. By leveraging existing active learning algorithms, we derive efficient edge evaluation policies which we apply on a spectrum of real world problems. We discuss insights from these preliminary results and potential research questions whose study may prove fruitful for both disciplines.