DisCo RL: Distribution-Conditioned Reinforcement Learning for General-Purpose Policies

TL;DR

DisCo RL introduces goal distributions as a flexible task representation for reinforcement learning, enabling the learning of general-purpose policies that generalize across diverse tasks, especially in robot manipulation.

Contribution

The paper proposes goal distributions for task representation and develops DisCo RL, an off-policy algorithm that effectively learns general-purpose policies with broad task generalization.

Findings

DisCo RL outperforms prior methods on robot manipulation tasks.

It effectively generalizes to new goal distributions.

The approach balances expressivity and learnability.

Abstract

Can we use reinforcement learning to learn general-purpose policies that can perform a wide range of different tasks, resulting in flexible and reusable skills? Contextual policies provide this capability in principle, but the representation of the context determines the degree of generalization and expressivity. Categorical contexts preclude generalization to entirely new tasks. Goal-conditioned policies may enable some generalization, but cannot capture all tasks that might be desired. In this paper, we propose goal distributions as a general and broadly applicable task representation suitable for contextual policies. Goal distributions are general in the sense that they can represent any state-based reward function when equipped with an appropriate distribution class, while the particular choice of distribution class allows us to trade off expressivity and learnability. We develop an off-policy algorithm called distribution-conditioned reinforcement learning (DisCo RL) to efficiently learn these policies. We evaluate DisCo RL on a variety of robot manipulation tasks and find that it significantly outperforms prior methods on tasks that require generalization to new goal distributions.