Researchers release Awaken-HI humanoid loco-manipulation model
Researchers have introduced Awaken-Human, a dataset pairing approximately 50,000 hours of egocentric stereo video with synchronized whole-body human motion, and Awaken-HI, a vision-language-action model for humanoid loco-manipulation. The model jointly predicts human motion and retargeted robot actions through a shared action expert, which researchers show improves skill transfer and reduces prediction error. After full-corpus pretraining and adaptation, Awaken-HI achieves a mean success rate of 93.75% across eight tasks and generalizes to unseen objects and destinations. Open article →