Task-Relevant Adversarial Imitation Learning

TL;DR

This paper identifies a vulnerability in adversarial imitation learning where discriminators focus on irrelevant features, and proposes TRAIL, a method that constrains discriminator learning to improve task performance in pixel-based robotic manipulation.

Contribution

The paper introduces TRAIL, a novel constrained discriminator approach that enhances the effectiveness of adversarial imitation learning for complex robotic tasks.

Findings

TRAIL outperforms standard GAIL and behavior cloning in robotic manipulation tasks.

Discriminator constraints improve task-relevant feature learning.

TRAIL successfully learns from pixels without explicit task rewards.

Abstract

We show that a critical vulnerability in adversarial imitation is the tendency of discriminator networks to learn spurious associations between visual features and expert labels. When the discriminator focuses on task-irrelevant features, it does not provide an informative reward signal, leading to poor task performance. We analyze this problem in detail and propose a solution that outperforms standard Generative Adversarial Imitation Learning (GAIL). Our proposed method, Task-Relevant Adversarial Imitation Learning (TRAIL), uses constrained discriminator optimization to learn informative rewards. In comprehensive experiments, we show that TRAIL can solve challenging robotic manipulation tasks from pixels by imitating human operators without access to any task rewards, and clearly outperforms comparable baseline imitation agents, including those trained via behaviour cloning and conventional GAIL.