LightParkour turns sparse human motion seeds into a deployable humanoid parkour policy

CyberRobooo · x · 2026-08-04

LightParkour turns sparse human motion clips into a deployable humanoid parkour policy

The paper shows that short human-motion seeds can be grounded in physics, expanded in simulation across terrain variations, and distilled into a single depth-based policy.

One policy handles both ordinary locomotion and contact-rich parkour on the Lightbot 0 humanoid.
Obstacles are treated as temporary supports, letting the robot intentionally use hands, knees, and body weight.
The method aims to avoid the usual tradeoff between motion-tracking policies that are scene-bound and reward-based locomotion that stays leg-dominant.
The authors argue this Real2Sim2Real pipeline can scale better than collecting a fresh demonstration for every obstacle or terrain.
They position the result as a step toward robots that can operate in cluttered, collapsed, or rescue-like environments without real-world fine-tuning.

Original post →

LightParkour turns sparse human motion seeds into a deployable humanoid parkour policy

More from Embodied