Relational Mimic for Visual Adversarial Imitation Learning

TL;DR

This paper presents Relational Mimic, a novel imitation learning method combining GANs and relational learning to improve robustness and sample efficiency in visual imitation tasks, with enhanced relational reasoning architectures.

Contribution

Introduces Relational Mimic, a new approach integrating relational learning with GANs for improved visual imitation learning and a novel neural network architecture for better relational reasoning.

Findings

Relational Mimic outperforms previous methods in pixel-based locomotion tasks.

Enhanced relational reasoning improves policy performance.

Ablation studies highlight the importance of relational learning components.

Abstract

In this work, we introduce a new method for imitation learning from video demonstrations. Our method, Relational Mimic (RM), improves on previous visual imitation learning methods by combining generative adversarial networks and relational learning. RM is flexible and can be used in conjunction with other recent advances in generative adversarial imitation learning to better address the need for more robust and sample-efficient approaches. In addition, we introduce a new neural network architecture that improves upon the previous state-of-the-art in reinforcement learning and illustrate how increasing the relational reasoning capabilities of the agent enables the latter to achieve increasingly higher performance in a challenging locomotion task with pixel inputs. Finally, we study the effects and contributions of relational learning in policy evaluation, policy improvement and reward learning through ablation studies.