Goal-Conditioned Reinforcement Learning with Imagined Subgoals

TL;DR

This paper introduces a goal-conditioned reinforcement learning method that uses imagined subgoals predicted by a high-level policy to improve learning of complex, temporally extended tasks, especially in robotics.

Contribution

It proposes a novel approach combining imagined subgoals with KL-constrained policy iteration to enhance learning efficiency and performance in complex tasks.

Findings

Outperforms existing methods on robotic navigation tasks

Improves learning speed and stability

Effectively handles temporally extended reasoning

Abstract

Goal-conditioned reinforcement learning endows an agent with a large variety of skills, but it often struggles to solve tasks that require more temporally extended reasoning. In this work, we propose to incorporate imagined subgoals into policy learning to facilitate learning of complex tasks. Imagined subgoals are predicted by a separate high-level policy, which is trained simultaneously with the policy and its critic. This high-level policy predicts intermediate states halfway to the goal using the value function as a reachability metric. We don't require the policy to reach these subgoals explicitly. Instead, we use them to define a prior policy, and incorporate this prior into a KL-constrained policy iteration scheme to speed up and regularize learning. Imagined subgoals are used during policy learning, but not during test time, where we only apply the learned policy. We evaluate our approach on complex robotic navigation and manipulation tasks and show that it outperforms existing methods by a large margin.