TeViR: Text-to-Video Reward with Diffusion Models for Efficient Reinforcement Learning

TL;DR

TeViR introduces a novel reward generation method using text-to-video diffusion models to improve sample efficiency and performance in reinforcement learning for robotic manipulation, outperforming traditional sparse reward approaches.

Contribution

The paper presents TeViR, a new approach that leverages pre-trained text-to-video diffusion models to generate dense rewards for RL, enhancing efficiency without ground truth rewards.

Findings

TeViR outperforms traditional sparse reward methods in 11 robotic tasks.

TeViR improves sample efficiency and overall performance.

TeViR works effectively without requiring ground truth environmental rewards.

Abstract

Developing scalable and generalizable reward engineering for reinforcement learning (RL) is crucial for creating general-purpose agents, especially in the challenging domain of robotic manipulation. While recent advances in reward engineering with Vision-Language Models (VLMs) have shown promise, their sparse reward nature significantly limits sample efficiency. This paper introduces TeViR, a novel method that leverages a pre-trained text-to-video diffusion model to generate dense rewards by comparing the predicted image sequence with current observations. Experimental results across 11 complex robotic tasks demonstrate that TeViR outperforms traditional methods leveraging sparse rewards and other state-of-the-art (SOTA) methods, achieving better sample efficiency and performance without ground truth environmental rewards. TeViR's ability to efficiently guide agents in complex environments highlights its potential to advance reinforcement learning applications in robotic manipulation.