Reproduce “Natural Humanoid Robot Locomotion with Generative Motion Prior” on OpenResearch

About this paper

Haodong Zhang, Liang Zhang, Zhenghan Chen, Lu Chen, Yue Wang, Rong Xiong

The Generative Motion Prior (GMP) framework enables humanoid robots to achieve natural, human-like locomotion by providing fine-grained, trajectory-level guidance from a pre-trained generative model. This approach demonstrates significantly higher motion naturalness, with a JFID of 0.931 compared to 2.088 for the best adversarial baseline, while maintaining robust velocity tracking on both simulated and physical robots.

Key insights

• Replacing scalar adversarial style rewards with stable, fine-grained, trajectory-level guidance from a generative model effectively overcomes instability and ambiguity in learning natural human-like movements.
• Offline pre-training and freezing a generative model creates a robust and detailed online supervision system for RL policies, enhancing training stability and interpretability.
• Achieving natural human-like locomotion through this method is crucial for improving human-robot interaction, social acceptance, and expanding the practical deployment of humanoid robots.

Summary by alphaXiv.