Amap, Alibaba’s mapping business, and Alibaba’s Token Hub team have released λ0 (lambda zero), a model that controls a humanoid robot’s whole body. Much of what it knows comes from people rather than robots: 500 hours of first-person video of humans going about everyday tasks. On the team’s own tests on a Unitree G1 robot, λ0 completed 62.5% of real-world trials, against 40% for π0.5, the strongest rival model it tested. None of these results has been reproduced by anyone outside the team.
The work is set out on a project page and in a preprint on arXiv, which has not been peer reviewed. The arXiv listing names the authors as the “Lambda-humanoid Team”. The name is a small joke: the GitHub repository says the shape of λ “echoes 人, the Chinese character for ‘human’.”
The model learns from people wearing cameras
Teaching robots usually means a person operating the robot by remote control, recording every move. That is slow and expensive. The λ0 team instead built a dataset it calls HumanVerse, which pairs “500 hours of robot-free capture across open-world scenes and tasks” with measurements of how the person’s body moved.
Each recording comes from a GoPro worn on the neck, set at a height that matches the robot’s own camera. Wearable PICO trackers follow 24 joints of the body, and software reconstructs 21 joints in each hand from the images. The team estimates this is about 3.4 times faster than collecting robot data. Averaging two kinds of task with equal weight, it puts a human recording at about 29 seconds against about 100 seconds for a robot one.
Training runs in three stages. First the model learns from seven existing collections of first-person human video. Then it trains on HumanVerse, which has been converted to match the robot’s body. Finally it is tuned for specific jobs using a mix of human demonstrations and recordings made on the robot. The human tracking is only for collecting data: “Human tracking is used during data collection, not at deployment,” the project page says. In use, the robot works from its head camera, a language instruction and its own sense of its joints.
The team’s results put λ0 ahead of four rival models
The test robot was a Unitree G1 fitted with BrainCo Revo 2 hands. The team ran four tasks (bottle disposal, toy storage, box transport and chair placement), with ten trials each. λ0 finished 25 of the 40 trials. The team reports these results for the other models:
- π0.5: 40.0% success
- StarVLA: 27.5%
- Ψ0: 10.0%
- GR00T N1.6: 7.5%
On a second measure, task progress, which gives partial credit for stages completed, λ0 averaged 80.4%. The paper says the people scoring the trials could not see which model was running. It also notes one task where the measures split: on chair placement λ0 succeeded six times to StarVLA’s three, but scored lower on progress.
In simulation, the team reports 90.0% success on SIMPLE, a public benchmark, across 180 trials. The paper adds a caveat: that figure “selects the best recorded configuration per task”, and “These numbers should not be interpreted as a fixed-configuration comparison.”
Removing the HumanVerse training stage cut real-world success to 22.5%, the team says, which is its main evidence that the human video matters.
The humanoid industry is betting on data and volume
The same wager is running across the humanoid business. Figure says its paid video-upload programme has gathered tens of millions of clips, as TSN reported when the Figure 04 reached design lock. Tesla says Optimus production has begun. China’s industry ministry expects humanoid output to pass 100,000 units this year. Robots in those numbers need software that can learn without a person steering each one by hand. Unitree, which makes the G1 used in the λ0 tests, already offers free UniUni software for directing groups of its humanoids at once.
The code is open, but the training data is not
The code is published under the permissive Apache 2.0 licence, and the team has put a mid-training checkpoint on Hugging Face, built on Alibaba’s Qwen3.5-2B model. The model card says it “is not a deployment-ready robot policy”. The README is equally frank. The HumanVerse and pretraining datasets are not included, the full final-stage training and robot deployment workflows are “Not yet released”, and the early-stage code is not guaranteed to reproduce the results end to end.
The self-comparison is stronger evidence than the league table
λ0 offers numbers for the idea that human video can stand in for robot data, and how they were measured is the interesting part. The cleanest number is the drop to 22.5% when the HumanVerse stage is taken out: that compares the team’s model with itself, under the same set-up, so it says more about the method than the headline score does. The comparison with π0.5 and the others is harder to weigh. Ten trials per task is a small sample, the paper reports no confidence intervals, and the rival models were run by the team on its own robot. A 62.5% success rate also means the robot failed on more than a third of attempts at everyday tasks. That is a research result, not a product.
Without HumanVerse, outsiders cannot yet test the central claim
The gaps in the release matter because the paper’s central claim is about data. Other labs can run the code, but without HumanVerse they cannot retrain the model the way the team did, so the figures above remain the team’s word for now. Owners of a G1 hoping for a ready-made brain face the same wait for the missing pieces. The test to watch is whether independent groups can match the 62.5% once the full recipe is out.
Sources
- λ₀: Scaling Open-World Egocentric Human Data for Humanoid Loco-Manipulation, project page
- Lambda-humanoid Team, arXiv 2610.18950
- LambdaHumanoid/openlambda on GitHub
- λ₀ Midtrain (100k) model card, Hugging Face
Related stories
- Figure Says Its Figure 04 Humanoid Has Reached Design Lock
- China Expects Humanoid Robot Output to Pass 100,000 This Year, MIIT Says
Follow @tsnmediahub on X for breaking AI news as it happens.
Follow on X
