Alibaba's Amap Open-Sources λ0 Humanoid, Trained on Human Video

Amap, Alibaba's mapping unit, and Alibaba Token Hub have open-sourced λ0, a humanoid foundation model trained entirely on first-person video of people doing everyday tasks rather than on robot teleoperation.

Transferred zero-shot to a Unitree G1, λ0 posts a 62.5% success rate across real-world locomotion-and-manipulation tasks, according to TechNode. On a smaller four-task, 40-trial comparison the team reports 25/40 (80.4%) for λ0 against 16/40 (54.0%) for the π0.5 baseline. Pull the HumanVerse mid-training stage out of the recipe and the Unitree G1's real-world score falls from 62.5% to 22.5%, which the team reads as evidence that the human video, and not just the architecture, is carrying the result.

HumanVerse itself is a 500-hour corpus of "32,993 episodes and 87 million pose frames" captured "in homes, shops, factories, offices and outdoors." Participants wore a GoPro HERO13, a Pico 4 Ultra headset and whole-body trackers; the paper notes "finger motion recovered from video afterward." The underlying stack is a Qwen3.5-2B vision-language backbone paired with a SONIC whole-body controller and BrainCo Revo2 hands. It is the 73rd Alibaba item we've logged in the past 90 days.

The GitHub release is Apache 2.0, but "HumanVerse-500 training data is not included in this code release." Only four motion-alignment examples and sample trajectories are public, and the release documents no participant consent or ethics review.