Alibaba's λ₀ humanoid model leverages 500h egocentric human data for zero-shot loco-manipulation
lambda-humanoid · hf · 2026-10-21
Alibaba's Token Hub and Amap team released λ₀ with the HumanVerse framework, training a humanoid foundation model for whole-body loco-manipulation from open-world egocentric human data.
Key points:
- Pairs egocentric video with synchronized body and hand motion, building a 500-hour open-world whole-body dataset
- Human-to-robot alignment pipeline enables zero-shot transfer to 7 real-world task categories at 62.2% success—no target-task robot demonstrations needed
- Reduces end-effector tracking error by 77.5% and orientation error by 69.9%
- Shows scaling trends between human data volume and task progress, validating egocentric data for humanoid training
More from Embodied
- Robotics insider: weak open-weight models make the field feel pre-ChatGPT — paigeinsf · 2026-10-21
- Agility Robotics' Digit heads to AGCO farm machinery plant in 2027 pilot — xmercury_one · 2026-10-21
- Google scientist visits humanoid robot fight event, poses with suspended robot — edgarpavlovsky · 2026-10-21
- Robot fight nights are becoming a tech-world group outing: RIK ring photos — edgarpavlovsky · 2026-10-21
- REK's new humanoid robots spotted on-site, human hangs alongside to mimic them — edgarpavlovsky · 2026-10-21
- 2028 ship date sparks speculation the robot is a prototype, platform to come — carlosdponx · 2026-10-21