Recently, AMAP, a subsidiary of Alibaba Group, officially released λ0, the first full-body foundation model based on open-world behavior. Relying solely on human action data, this model enables robots to learn full-body coordination of locomotion, posture, and bimanual manipulation, completely freeing robots from the limitation of “partial action learning”. On the SIMPLE simulation benchmark and multiple real-robot evaluations, λ0 achieved SOTA performance, with data collection efficiency approximately 3.4 times that of mainstream methods, and a real-robot task success rate 22.5% higher than the strongest comparison baseline.
As embodied intelligence gradually moves into open environments such as homes, supermarkets, and warehouses, robots no longer face single operations under fixed standing postures, but rather complex long-horizon tasks that require continuous coordination of full-body movements.
However, existing technical approaches struggle to support the demands of full-body coordinated data collection and learning for robots. On the one hand, traditional real-robot teleoperation collection is highly dependent on hardware equipment and manual resetting, making scaling extremely costly; on the other hand, previous first-person-view VLA (Vision-Language-Action) foundation models generally lack full-body coordinated action information, with data often focusing only on local interactions such as hands and arms, and lacking whole-body dynamic information such as joints, torso, and pose. As scenarios become more open and tasks more complex, this “fragmented” action learning becomes a bottleneck restricting robots’ general capabilities and action success rates.
λ0 pioneered the technical approach of “body-free full-body data collection + hardware-software collaborative human-machine alignment” to address the challenges of collection efficiency, full-body coordination, and cross-embodiment transfer in open scenarios.
In terms of cost and efficiency, body-free full-body data collection frees λ0 from heavy dependence on expensive robot-embodied collection; collectors only need to wear consumer-grade devices to collect data on the go in everyday environments. Compared with robot-embodied collection, this method reduces the collection time per demonstration for representative tasks from about 100 seconds to 29 seconds, improving collection efficiency by approximately 3.4 times.
In terms of full-body action learning and alignment, λ0 relies on a hardware-software collaborative data collection mode to fully record and align full-body action information such as vision, body displacement, and hand manipulation, and then, through three-stage training covering hand interaction, full-body coordination, and humanoid embodiment, gradually converts human experience into full-body control commands for robots. Based on this scalable pipeline, AMAP built the HumanVerse-500 dataset containing 500 hours of open-world body-free full-body data.
(Construction and panoramic statistics of the HumanVerse-500 dataset)
Facing the transfer challenges posed by human-robot morphological differences and end-effector execution deviations, λ0 designs a viewpoint-matching and “coarse-to-fine” action alignment mechanism at the collection, processing, and action ends, optimizing physical contact precision while preserving full-body coordination.
In terms of cross-morphology transfer and data scalability (Scaling), λ0 achieves zero-shot transfer on real robots across 7 full-body coordination and long-horizon mobile manipulation tasks using only human action data. Meanwhile, under the same amount of robot demonstrations, as the scale of human full-body data expands, the real robot's task performance improves significantly. This means that scaling up the accumulation of human full-body data is an effective path to improving the real-world task performance of humanoid robots.
(Task progress of aligning with human demonstrations across the seven tasks)
It is reported that this model research was jointly completed by AMAP and Alibaba's ATH business group. From data collection that requires no robot presence, to human-machine observation and action alignment, and then to capability verification on real robots, λ0 has opened up the complete pipeline for bringing human full-body experience to humanoid robots. This provides a foundation for robots to learn and execute full-body coordinated tasks in more open scenarios such as inspection and logistics.
Currently, λ0 has open-sourced the relevant hardware configuration code on GitHub and released HumanVerse data samples on HuggingFace, making it convenient for developers to directly experience and conduct research.
Appendix
Project homepage: https://lambdahumanoid.github.io/
Technical report: https://arxiv.org/abs/2610.18950
GitHub: https://github.com/LambdaHumanoid/openlambda
HumanVerse data samples: https://huggingface.co/datasets/lambda-humanoid/HumanVerse-500-sample
This article was provided by AMAP and reprinted by QbitAI with authorization; the views belong to the original author.