Training Transition Policies via Distribution Matching for Complex Tasks

TL;DR

This paper introduces transition policies trained via adversarial inverse reinforcement learning to smoothly connect lower-level policies in hierarchical reinforcement learning, improving success rates on complex tasks by matching state-action distributions.

Contribution

It proposes a novel method for training transition policies using distribution matching and deep Q-learning, addressing the challenge of sparse rewards in hierarchical RL.

Findings

Achieves higher success rates in continuous locomotion and manipulation tasks.

Effectively matches state-action distributions between policies.

Outperforms previous trajectory search methods.

Abstract

Humans decompose novel complex tasks into simpler ones to exploit previously learned skills. Analogously, hierarchical reinforcement learning seeks to leverage lower-level policies for simple tasks to solve complex ones. However, because each lower-level policy induces a different distribution of states, transitioning from one lower-level policy to another may fail due to an unexpected starting state. We introduce transition policies that smoothly connect lower-level policies by producing a distribution of states and actions that matches what is expected by the next policy. Training transition policies is challenging because the natural reward signal -- whether the next policy can execute its subtask successfully -- is sparse. By training transition policies via adversarial inverse reinforcement learning to match the distribution of expected states and actions, we avoid relying on task-based reward. To further improve performance, we use deep Q-learning with a binary action space to determine when to switch from a transition policy to the next pre-trained policy, using the success or failure of the next subtask as the reward. Although the reward is still sparse, the problem is less severe due to the simple binary action space. We demonstrate our method on continuous bipedal locomotion and arm manipulation tasks that require diverse skills. We show that it smoothly connects the lower-level policies, achieving higher success rates than previous methods that search for successful trajectories based on a reward function, but do not match the state distribution.