SPRINT: Efficient Spectral Priors for Humanoid Athletic Sprints

Researchers introduced SPRINT, a frequency-adaptive spectral prior framework for humanoid sprinting.

Using a small motion library, it learns frequency-domain priors to generate stable, kinematically feasible trajectories.

The method achieved zero-shot sim-to-real transfer on the Unitree G1 humanoid robot.

In real-world tests, the robot sprinted at up to 6 m/s, highlighting a data-efficient path to high-speed locomotion.