LightParkour grows humanoid parkour skills via Real2Sim2Real into one deployable depth policy

chris_j_paxton · x · 2026-09-11

LightParkour uses humanoid parkour as a testbed for contact-rich skill learning, and the RoboPapers podcast will soon release a deep-dive episode with the authors.

Core method:

Compared to prior routes: motion-tracking policies are expressive but tied to demonstrated scenes, while reward-driven locomotion adapts to terrain yet skews leg-dominant; LightParkour achieves geometry-aware, load-bearing contact.

Original post →

More from Embodied

Embodied channel →