Exploring Humanoid Robot Training: Learning Unseen Terrain via Synthetic Routes

wandb · x · 2026-08-12

A developer shared the Day 251 progress on humanoid robot training. They are testing a hypothesis: can an agent that has learned base physical skills from real clips continue to learn unseen varied terrain using only sparse synthetic routes and observation/reward magic?

If successful, this approach will massively reduce the real motion samples required to teach humanoid skills. Despite encountering a 90-degree mismatch bug between pelvis keyframes and joystick controls, the latest logs show the system is surprisingly robust with excellent recovery capabilities during synthetic episodes.

Original post →

More from Embodied

Embodied channel →