DisCo RL: Distribution-Conditioned Reinforcement Learning for General-Purpose Policies
This addresses the challenge of flexible and reusable skills in reinforcement learning for robotics, representing an incremental advance over existing contextual and goal-conditioned methods.
The paper tackled the problem of learning general-purpose policies for a wide range of tasks by proposing goal distributions as a task representation, and found that their DisCo RL algorithm significantly outperformed prior methods on robot manipulation tasks requiring generalization to new goal distributions.
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.