Alan Hou

Latest article / 最新文章

DreamX: Repurposing Video Diffusion Models as Universal Latent Action Learners

DreamX：将视频扩散模型重构为通用隐式动作学习器

Translating human internet videos into robot manipulation policies has long been hindered by embodiment mismatch and appearance entanglement. KAIST and NVIDIA researchers present DreamX, which repurposes a pretrained video diffusion model to extract compact latent actions directly from spatiotemporal attention maps—compressing video frames by 300x (1,200 to 4 tokens). DreamX achieves state-of-the-art representation quality on LARY-Bench, enables zero-shot human-to-robot action transfer, and slashes video policy compute by over 80%.