Reproduce “Humanoid Parkour Learning” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ziwen Zhuang, Shenzhe Yao, Hang Zhao

A humanoid robot policy was developed that allows the Unitree H1 to perform a diverse range of parkour skills, including jumping and leaping over gaps, using only onboard vision. The system achieves robust zero-shot sim-to-real transfer without relying on explicit motion priors or complex reward engineering.

Key insights

  • Training with fractal noise terrain effectively replaces explicit "feet air time" rewards or motion references, leading to more diverse and robust emergent gaits for humanoids.
  • Onboard depth vision is crucial for enabling agile parkour skills, allowing the humanoid to autonomously perceive and adapt to diverse obstacles.
  • Efficient sim-to-real transfer for complex whole-body control policies benefits significantly from initializing the student policy with an expert (oracle) policy and leveraging multi-GPU accelerated distillation.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.