Amap, Alibaba Group's mapping and location unit, has released Lambda-0, which it describes as the first whole-body foundation model for humanoid robots built on open-world human behavior. The research was carried out with Alibaba's ATH business group, and the code is on GitHub under an Apache 2.0 license alongside an arXiv paper.
The pitch is that a humanoid can learn walking, posture and two-handed manipulation together without a robot present during data collection. Many first-person vision-language-action models focus on hands and arms, the team argues, and miss the joint, torso and body-pose information that long, open-environment tasks require. Teleoperating real robots to fill that gap is expensive and slow to scale.
Lambda-0 instead uses what Amap calls robot-free whole-body data collection. Collectors wear consumer-grade devices and record as they move through everyday spaces. For a representative task, a single demonstration took about 29 seconds to collect, compared with roughly 100 seconds with a physical robot, about 3.4 times the efficiency. Using this pipeline, the team built HumanVerse-500, a dataset of 500 hours of open-world whole-body human data that aligns vision, body movement and hand actions.
Training runs in three stages: hand interaction, whole-body coordination and finally the humanoid embodiment. To cope with the differences between human and robot bodies and with end-effector errors, the model uses viewpoint matching and a coarse-to-fine action-alignment scheme that keeps whole-body coordination while improving contact precision.
According to the paper, Lambda-0 achieved state-of-the-art results on the SIMPLE simulation benchmark, where it reports about 90 percent, and in real-robot tests. On seven whole-body and long-horizon mobile manipulation tasks, it transferred zero-shot to a Unitree G1 using only human data, and the team reports a real-robot success rate of 62.5 percent, 22.5 percentage points above the strongest baseline. With robot demonstrations held constant, adding more human whole-body data kept improving real-robot performance, which the team presents as evidence that human data can scale humanoid learning.
The figures come from the team's paper and have not yet been independently reproduced. For Amap, the release adds an embodied track to its recent world-model work on 3D cities.
SEE ALSO: Amap Releases ABot-Earth 0.7 · Peking University and PrimeBot's HIL-UMI