O2LaF repurposes video diffusion transformers for SOTA self-supervised latent action learning

NationalUniversityofSingapore · hf · 2026-11-20

Researchers from NUS Show Lab, Tencent, and Stanford present O2LaF, which reuses a single pretrained video diffusion transformer to learn inverse and forward latent action dynamics from unlabeled videos. Its unified latent action space spans first/third-person game footage and robot manipulation, achieving state-of-the-art few-shot action probing and zero-shot action transfer without a dedicated robot-specific action encoder.

Original post →

More from Embodied

Embodied channel →