λ_0: Scaling Open-World Egocentric Human Data for Humanoid Loco-Manipulation

λ_0. Scaling Open-World Egocentric Human Data for Humanoid Loco-Manipulation. By humanoid Team, from Alibaba Token Hub and Amap, Alibaba Group. Published on the thirteenth of October. Its code repository is worked out: about one hundred source files, with a readme and a licence.

Humanoid loco-manipulation means moving and handling objects at the same time, with the whole body coordinated against the physical world. Robot demonstrations of it are costly to collect at scale. Human demonstrations are the scalable alternative, but existing egocentric human data usually leaves out locomotion and posture. The new resource is HumanVerse, a dataset built to fill that hole.

HumanVerse holds five hundred hours of open-world human loco-manipulation. Each recording pairs egocentric video with synchronized motion of the body and the hands, so what the person saw and how their whole body moved arrive together.

Human data is not robot data, and two things stand in the way: differences in embodiment, and controller dynamics that make the end effector land somewhere other than intended. The human-to-humanoid pipeline addresses both. It aligns camera viewpoints and the appearance of the arm, and offline it refines the motion references and controller inputs to cut end-effector error while keeping whole-body coordination intact. The pipeline is validated by zero-shot transfer to a Unitree G1, on representative whole-body manipulation and long-range loco-manipulation tasks, with no robot demonstrations of the target tasks.

The first user is the authors' own model, lambda zero, a humanoid foundation model for whole-body loco-manipulation. Robot-aligned human demonstrations from this pipeline supply its mid-training supervision. Adding more whole-body human data consistently improves downstream task progress. Lambda zero reaches ninety percent success on SIMPLE, and sixty-two point five percent across real-world loco-manipulation tasks.

The scaling result is the one to keep. More human data, made to look and move like robot data, kept making the robot better, which is the property any cheaper data source has to show before it can stand in for the expensive kind. The authors offer the study as a guide for collecting whole-body human data at scale, aimed at humanoid foundation models that interact more richly with the physical world.