Reproduce “Learning Generic and Dynamic Locomotion of Humanoids Across Discrete Terrains” on OpenResearch

About this paper

Shangqun Yu, Nisal Perera, Daniel Marew, Donghyun Kim

A hierarchical control architecture enables humanoid robots to learn generic and dynamic locomotion behaviors, including walking, jumping, and leaping, across discrete terrains. This approach leverages a reinforcement learning policy trained efficiently in a simplified environment, achieving 1-2 orders of magnitude faster convergence than end-to-end reinforcement learning and demonstrating zero-shot transferability across different humanoid platforms.

Key insights

• Decoupling high-level strategic decision-making in a simplified, abstract environment from low-level dynamic execution significantly improves reinforcement learning training efficiency and policy generalizability.
• Integrating perception-driven gait selection (walking, leaping, jumping) with optimal motion control enables humanoids to robustly navigate complex discrete terrains.
• A null-space projection-based whole-body controller allows for seamless integration of auxiliary manipulation tasks without interfering with learned locomotion or requiring policy retraining.

Summary by alphaXiv.