Lambda-0

Lambda-0 (λ₀) is a whole-body humanoid foundation model released and open-sourced by Alibaba's Amap (mapping unit) together with Alibaba Token Hub (ATH). Its pitch: learn walking and manipulation together (loco-manipulation) from first-person human data only, with no robot present during data collection.

The paper is arXiv 2610.18950 and the code is on GitHub under Apache 2.0; Pandaily covered the open-source release on October 24, 2026. The λ resembles the Chinese character 人 ("human").

The idea

Humanoid training usually relies on costly, slow real-robot teleoperation, and most egocentric VLA models focus on hands and arms while ignoring torso, posture and locomotion. Lambda-0 uses robot-free whole-body capture: collectors wear a neck-mounted GoPro and PICO body trackers and simply do tasks in everyday spaces.

HumanVerse-500 dataset

• 500 hours of open-world human loco-manipulation: 32,993 episodes, 87M pose frames, 827 task types
• Egocentric RGB + 24 tracked body joints + 21 joints per hand (reconstructed from video), all synchronized
• ~29 s per demo vs ~100 s on a real robot, about 3.4× collection throughput

Three-stage training

1. Egocentric hand-motion pretraining on seven human datasets
2. Robot-aligned whole-body human mid-training: human demos are converted into the robot's observation/action space (viewpoint matching, rendered G1 arms, coarse-to-fine action alignment)
3. Robot post-training with task-aligned human demos plus native robot demos

Results (from the paper, not yet independently reproduced)

• On a Unitree G1 (BrainCo Revo2 hands, SONIC whole-body controller): 62.5% real-world success (25/40), +22.5 pts over the strongest baseline π₀.₅ (40%); 80.4% task progress
• 90.0% success on the SIMPLE simulation benchmark
• More human whole-body data keeps improving real-robot performance

How to use / follow it

• Project page: lambdahumanoid.github.io (videos and dataset overview)
• Paper: arXiv 2610.18950, "λ₀: Scaling Open-World Egocentric Human Data for Humanoid Loco-Manipulation"
• Code: open source on GitHub, Apache 2.0
• At deployment: the policy takes the robot head-camera image, a language instruction and proprioception, and predicts whole-body motion tokens plus hand commands; SONIC executes the body motion. Human tracking is only used during data collection.

Who it's for

• Researchers working on humanoid whole-body control, VLAs and human-to-robot transfer
• Embodied-AI teams looking for cheaper ways to scale robot training data

Related work: EgoHumanoid-V2, EgoAlign. Check the official repository for current code and weights.