Event

Research on Pretraining HuGe Humanoid General-Purpose VLA Models

1 reports · 1 sources

Yang Liu and other authors proposed and pretrained the HuGe series of humanoid general-purpose vision-language-action (VLA) models, using over 2,000 hours of first-person human and robot data. The models follow three existing representative VLA architectures, predict latent motion tokens to drive a pretrained whole-body controller, and output hand control signals. To address the mismatch between human data and the robot embodiment, the authors construct pseudo-proprioceptive states via simulation replay during pretraining, and modify the arm appearance in first-person observations via re-rendering during downstream fine-tuning. The authors report that real-world experiments on the Unitree G1 show that augmenting robot demonstrations with visually adapted first-person demonstrations significantly improves the success rate of pick-and-place tasks.