KAIST AI Develops DreamX for 300x Video Diffusion Model Compression in Robotics

KAIST AI has introduced DreamX, a system that significantly compresses pretrained video diffusion models by 300 times for robotics applications. This innovation allows for more compute-efficient robot policies by extracting action-relevant dynamics and reusing diffusion priors. DreamX demonstrates state-of-the-art performance and zero-shot cross-embodiment transfer on LARY-Bench, advancing efficient robot action representation.