Versatile Skill Control via Self-supervised Adversarial Imitation of Unlabeled Mixed Motions

TL;DR

This paper introduces a self-supervised adversarial imitation learning method that learns versatile, controllable skills from unlabeled mixed motion data, enabling robots to acquire diverse behaviors without explicit labels.

Contribution

It proposes a cooperative adversarial approach for learning controllable, versatile policies from unlabeled datasets, and demonstrates emergent useful skills through unsupervised discovery within a generative adversarial framework.

Findings

Successful skill replication on Solo 8 robot

Emergence of useful skills without labeled data

Versatile policies outperform prior methods

Abstract

Learning diverse skills is one of the main challenges in robotics. To this end, imitation learning approaches have achieved impressive results. These methods require explicitly labeled datasets or assume consistent skill execution to enable learning and active control of individual behaviors, which limits their applicability. In this work, we propose a cooperative adversarial method for obtaining single versatile policies with controllable skill sets from unlabeled datasets containing diverse state transition patterns by maximizing their discriminability. Moreover, we show that by utilizing unsupervised skill discovery in the generative adversarial imitation learning framework, novel and useful skills emerge with successful task fulfillment. Finally, the obtained versatile policies are tested on an agile quadruped robot called Solo 8 and present faithful replications of diverse skills encoded in the demonstrations.