Alibaba's λ0 trains a humanoid on 500 hours of GoPro human video, 62.5% vs π0.5's 40%
AIFlow_ML · x · 2026-10-26
Alibaba's λ0 (HumanVerse) project trains a Unitree G1 humanoid on 500 hours of first-person human video recorded with neck-mounted GoPros.
- Method: pairs human activity video with whole-body motion, using alignment plus three-stage training for zero-shot transfer to robot control.
- Results: 62.5% success on box-carrying and toy-tidying tasks vs 40% for π0.5 in the team's own tests; 76.2% progress on EgoAlign long-range loco-manipulation and 51.7% on whole-body coordination tasks.
- Demos: real lab/break-room/visitor-center scenes — boxes, fridges, watering plants, pedal bins, moving targets.
- Caveat: the authors note the more convincing number is the ablation without human data; scores are self-reported.
- Project page, preprint, code, and HF data are open.
Related event: Alibaba's λ0 Trains Humanoid Robot with 500 Hours of Human Video(4 posts)→
More from Embodied
- On-device VLM goes production: GenieX runs Liquid AI's LFM-VL live at PyTorch Conference — PyTorch · 2026-10-26
- LeCun co-authored Temporal Straightening: straightening latent paths boosts gradient-based planning success rates — ylecun · 2026-10-26
- Humanoid robots' DRAM demand may 20x, but it's still just 0.04% of 2026 supply — zephyr_z9 · 2026-10-26
- Microduck team publishes deep dives: relocalizes within 2cm using a 64-pixel depth sensor — mishig25 · 2026-10-26
- Robot's elbow-assisted get-up move goes viral for its humanlike fluidity — iamfakhrealam · 2026-10-26
- Chinese eVTOL EHang EH216-S completes nearly 100,000 flights across 23 countries — LoganGrasby · 2026-10-26