Alibaba's λ0 humanoid learns from GoPro videos, hits 62.5% on Unitree G1 vs π0.5's 40%
AIFlow_ML · x · 2026-10-26
Alibaba's λ0 humanoid approach trains robots like box-carrying and toy-tidying mostly from human videos captured with a neck-mounted GoPro.
- Trained on roughly 500 hours of human video
- Reported 62.5% success on a Unitree G1 in the team's own tests, versus 40% for π0.5
- The poster hints "the number that matters is a different one", suggesting a more significant metric beyond these self-reported results
Related event: Alibaba's λ0 Trains Humanoid Robot with 500 Hours of Human Video(4 posts)→
More from Embodied
- RoboJudge: 800 human-annotated videos benchmark MLLM judges for embodied video generation — CUHK-CSE · 2026-10-26
- HKUST transfers Unitree G1 controller to full-size H2 with one GPU in ~8 hours — FinanceYF5 · 2026-10-26
- China Mobile shows non-invasive BCI + AR glasses controlling a robot hand — FinanceYF5 · 2026-10-26
- Berkeley diffusion model generates RL policy weights, letting a real G1 dance — FinanceYF5 · 2026-10-26
- KAIST Humanoid v0.9 pushes 35 kg and lifts 20 kg with onboard vision — FinanceYF5 · 2026-10-26
- Figure F.03 logs 422 miles across 13 days — and the treadmill test continues — FinanceYF5 · 2026-10-26