Reproduce “SPRINT: Efficient Spectral Priors for Humanoid Athletic Sprints” on OpenResearch

About this paper

Yantong Wei, Kaihong Huang, Hainan Pan, Jiawei Luo, Jiawei Zhou, Ziyan Mai, Zhiwen Zeng, Yaonan Wang and 1 more

The SPRINT framework introduces frequency-adaptive spectral priors and a hierarchical residual control architecture to enable stable, biomimetic athletic sprints for humanoid robots, achieving speeds up to 6 m/s on a physical Unitree G1 robot with zero-shot sim-to-real transfer. It utilizes a data-efficient approach, deriving kinematic guidance from a small set of human motion sequences to achieve continuous gait transitions.

Key insights

• Frequency-adaptive spectral priors can effectively derive fine-grained kinematic guidance from a minimal set of human motion data, enabling robust extrapolation of joint trajectories beyond reference speeds.
• A hierarchical control architecture combining spectral priors for biomimicry and residual reinforcement learning for stabilization yields stable, high-speed, and natural humanoid locomotion.
• The combination of asymmetric actor-critic, curriculum learning, and dynamics randomization is crucial for achieving successful zero-shot transfer of complex, dynamic gaits from simulation to physical robots.

Summary by alphaXiv.