Collecting robot demonstrations for humanoid loco-manipulation is expensive, so researchers often turn to human videos. But existing egocentric datasets typically omit locomotion and posture, and mismatches between human and robot bodies make it hard to transfer precise end-effector movements. The Lambda-humanoid Team introduces HumanVerse, a 500-hour dataset of open-world human loco-manipulation that pairs egocentric video with synchronized body and hand motion, explicitly capturing whole-body coordination.
To bridge the embodiment gap, the team built a pipeline that aligns camera viewpoints and arm appearance between human and robot, then refines motion references and controller inputs offline. This reduces end-effector errors while preserving whole-body coordination. They validate the approach through zero-shot transfer to a Unitree G1 humanoid on both whole-body manipulation and long-range loco-manipulation tasks, with no target-task robot demonstrations required.
Using robot-aligned human demonstrations for mid-training, they developed λ0, a humanoid foundation model for whole-body loco-manipulation. Scaling up whole-body human data consistently improved downstream task progress. λ0 achieved 90.0% success on the SIMPLE benchmark and 62.5% across real-world loco-manipulation tasks, suggesting that large-scale egocentric human data can be a practical path toward humanoid foundation models for richer physical interaction.