Diffusion Reward is a novel framework that learns rewards from expert videos for reinforcement learning tasks. The system leverages conditional video diffusion models to solve complex visual reinforcement learning problems. The primary insight is that lower generative diversity is noted when conditioned on expert trajectories. Diffusion Reward encourages the exploration of expert-like behaviors and has proven effective over numerous robotic manipulation tasks. The system can also solve unseen tasks and outperforms baseline methods.
Publication date: 21 Dec 2023
Project Page: diffusion-reward.github.io
Paper: https://arxiv.org/pdf/2312.14134
