Transferring Agent Behaviors from Videos via Motion GANs

TL;DR

This paper presents a method using Motion GANs to generate meaningful sub-goals from raw video pixels, facilitating the training of reinforcement learning agents in new environments.

Contribution

It introduces a novel approach that automatically derives sub-goals from videos using Motion GANs, aiding reinforcement learning without manual reward design.

Findings

Generated behaviors are visually meaningful in unseen environments.

Motion templates effectively serve as sub-goals for RL training.

The approach reduces the need for manual reward specification.

Abstract

A major bottleneck for developing general reinforcement learning agents is determining rewards that will yield desirable behaviors under various circumstances. We introduce a general mechanism for automatically specifying meaningful behaviors from raw pixels. In particular, we train a generative adversarial network to produce short sub-goals represented through motion templates. We demonstrate that this approach generates visually meaningful behaviors in unknown environments with novel agents and describe how these motions can be used to train reinforcement learning agents.