Motion learning in variable environments using probabilistic flow tubes

Abstract

Commanding an autonomous system through complex motions at a low level can be tedious or impractical for systems with many degrees of freedom. Allowing an operator to demonstrate the desired motions directly can often enable more intuitive and efficient interaction. Two challenges in the field of learning from demonstration include (1) how to best represent learned motions to accurately reflect a human's intentions, and (2) how to enable learned motions to be easily applicable in new situations. This paper introduces a novel representation of continuous actions called probabilistic flow tubes that can provide flexibility during execution while robustly encoding a human's intended motions. Our approach also automatically determines certain qualitative characteristics of a motion so that these characteristics can be preserved when autonomously executing the motion in a new situation. We demonstrate the effectiveness of our motion learning approach both in a simulated two-dimensional environment and on the All-Terrain Hex-Limbed Extra-Terrestrial Explorer (ATHLETE) robot performing object manipulation tasks.