Hang Zhao

Co-training generative world models on 1,880 hours of retargeted human motion and video unlocks an 84% real-world humanoid loco-manipulation success rate while drastically slashing the need for costly teleoperated robot data.

Purely autoregressive VLAs no longer require awkward hybrid diffusion heads to achieve fine-grained robot control: structuring discrete action tokenization as causally annealed flow matching recovers continuous diffusion precision while preserving end-to-end token efficiency.

This work proposes an observer-based dynamic cooperative learning framework incorporating online distributed Gaussian Process (GP) regression, which enables accurate estimation despite incomplete in measurements and deficient GP models, and derives an error upper bound encompassing state estimation and model estimation.