Amap's λ0 trains humanoid whole-body control from human data alone, with 3.4x faster collection

Amap, a unit of Alibaba Group, has released λ0, which it describes as the first whole-body foundation model built on open-world human behavior. According to Amap, the model teaches robots whole-body coordination (locomotion, posture and bimanual manipulation) using only human action data. The announcement was published in Chinese by Qbitai (量子位) on 2026-10-23 and republished there with Amap's authorization.

Key figures (Amap's claims)

• Benchmarks: Amap says λ0 reached state-of-the-art results on the SIMPLE simulation benchmark and in several real-robot evaluations. Specific benchmark scores are not given in the announcement.
• Real-robot success: 22.5% higher task success than the strongest comparison baseline.
• Collection efficiency: about 3.4 times that of mainstream methods. For a representative task, one demonstration takes 29 seconds to collect, down from about 100 seconds with robot-body collection.
• Dataset: HumanVerse-500, containing 500 hours of open-world whole-body data collected without a robot body.
• Transfer: zero-shot transfer to real robots on 7 whole-body coordination and long-horizon mobile manipulation tasks, using only human action data.

How it works

Amap says λ0 uses "body-free whole-body data collection" combined with hardware-software co-designed human-robot alignment. Collectors wear consumer-grade devices and record data while moving through everyday environments, with no robot present and no manual robot resets.

The collection setup records and aligns visual input, body displacement and hand manipulation. Training then runs in three stages: hand interaction, whole-body coordination and humanoid embodiment. Amap says this converts human experience into whole-body robot control commands.

To handle differences between human and robot morphology and end-effector execution error, the model uses viewpoint matching and a coarse-to-fine action alignment mechanism. Amap says this preserves whole-body coordination while improving physical contact precision.

Amap also reports that, with robot demonstration volume held constant, task performance on real robots improves markedly as the amount of human whole-body data grows.

Problem addressed

Amap argues that two limits block whole-body learning. Teleoperated real-robot collection depends on costly hardware and manual resets. Earlier egocentric vision-language-action (VLA) foundation models generally lack whole-body motion information such as joints, torso and pose, and focus on hands and arms.

Release and availability

The research was carried out by Amap and Alibaba's ATH business group. Amap has open-sourced hardware configuration code on GitHub and published a HumanVerse sample dataset on Hugging Face.

• Project page: lambdahumanoid.github.io
• Technical report: arxiv.org/abs/2610.18950
• Code: github.com/LambdaHumanoid/openlambda
• Data sample: huggingface.co/datasets/lambda-humanoid/HumanVerse-500-sample

What this means

λ0 targets the data bottleneck in humanoid learning. Amap's reported numbers (29 seconds versus about 100 seconds per demonstration, and zero-shot transfer on 7 tasks) point to human wearable-device collection as a cheaper substitute for robot-body teleoperation. The scaling result Amap reports, better real-robot performance as human whole-body data grows, supports its argument that accumulating human data is an effective path for humanoid task performance. All results are Amap's own, and independent replication will depend on the released code, data sample and technical report. Amap says the pipeline provides a foundation for whole-body tasks in settings such as inspection and logistics.