Lambda-0
Amap and Alibaba's ATH group released Lambda-0, a whole-body humanoid model trained on 500 hours of human first-person data, which transfers zero-shot to a Unitree G1 robot.
Earlier first-person vision-language-action models mostly focus on hands and arms. Lambda-0 also learns walking, posture and two-handed manipulation from human data collected with consumer devices and no robot present. Training runs in three stages (hand interaction, whole-body coordination, humanoid embodiment), with viewpoint matching and coarse-to-fine action alignment to bridge human and robot bodies.
- Date
- Saturday 24 October 2026
- Lab
- Amap (Alibaba ATH)
- Kind
- open-weights
- Access
- open weights
Figures
| Measure | Value | Measured by |
|---|---|---|
| Real-robot success rate on seven whole-body and long-horizon tasks (Unitree G1, zero-shot) | 62.5% 22.5 percentage points above the strongest baseline | company |
| SIMPLE simulation benchmark | about 90% reported in the paper, not independently reproduced | company |
| HumanVerse-500 dataset size | 500 hours open-world whole-body human data | company |
| Time to collect one demonstration | about 29 s vs about 100 s with a physical robot about 3.4 times the efficiency | company |
Code is on GitHub under Apache 2.0 with an arXiv paper. The page does not say whether weights are released, and the figures have not been independently reproduced.
Sources
This record was partly confirmed: some claims could not be checked on 24 October 2026. How we check