Direct then Diffuse: Incremental Unsupervised Skill Discovery for State Covering and Goal Reaching

TL;DR

This paper introduces UPSIDE, a method for unsupervised skill discovery that balances environment coverage with directed goal-reaching, improving the learning of diverse, discriminable behaviors in reinforcement learning without rewards.

Contribution

UPSIDE proposes a novel decoupled policy structure and a coverage-directed optimization framework, advancing unsupervised skill discovery by effectively balancing coverage and directedness.

Findings

UPSIDE outperforms existing methods in navigation tasks.

Learned skills enable better downstream task performance.

The method effectively balances coverage and goal-directed behavior.

Abstract

Learning meaningful behaviors in the absence of reward is a difficult problem in reinforcement learning. A desirable and challenging unsupervised objective is to learn a set of diverse skills that provide a thorough coverage of the state space while being directed, i.e., reliably reaching distinct regions of the environment. In this paper, we build on the mutual information framework for skill discovery and introduce UPSIDE, which addresses the coverage-directedness trade-off in the following ways: 1) We design policies with a decoupled structure of a directed skill, trained to reach a specific region, followed by a diffusing part that induces a local coverage. 2) We optimize policies by maximizing their number under the constraint that each of them reaches distinct regions of the environment (i.e., they are sufficiently discriminable) and prove that this serves as a lower bound to the original mutual information objective. 3) Finally, we compose the learned directed skills into a growing tree that adaptively covers the environment. We illustrate in several navigation and control environments how the skills learned by UPSIDE solve sparse-reward downstream tasks better than existing baselines.