DreamX preprint reports 300× compression of video action representations

Researchers report in a DreamX preprint that they compressed each latent video frame from 1,200 tokens to 4 while retaining action information that transfers across human and robot embodiments and environments. The method repurposes a pretrained video diffusion model, extracting compact representations of actions from features that capture movement over space and time. The authors report leading performance among latent action models on LARY-Bench and action transfer without additional training across different embodiments and environments. The shared representation also supports policy learning in a compact action space, reducing the computational cost of video-diffusion-based policies. The findings are reported in a preprint.