Towards a General Humanoid Loco-Manipulation Model via Egocentric Whole-Body Human Data Pretraining
The research developed HUMANVERSE-500, a 500-hour dataset of diverse human locomanipulation behaviors, and a unified policy called λ0. This policy uses a three-stage training recipe—interaction pre-training, whole-body data mid-training, and embodiment post-training—to transfer human experience to robot control. The method achieves state-of-the-art performance in real-world manipulation tasks.