HuGe Scaling Humanoid VLAs via 2,000 Hours of Egocentric Data
October 24, 2026
HuGe scales humanoid Vision-Language-Action models by pre-training on 2,000 hours of egocentric human and robot data. The model predicts latent motion tokens for a pretrained whole-body controller, enabling transfer from human demonstrations to humanoid embodiments.
HOW THIS AFFECTS YOU
●
builderYou can leverage large-scale egocentric datasets to train humanoid policies that require coordinated whole-body control.
●
researcherThe three-stage training pipeline addresses the visual and proprioceptive gaps in human-to-robot transfer.