RynnVLA-Latent trains 4B robot policy on 30.1K hours of video

A 4B-parameter robot policy claims 98.8% success on the LIBERO benchmark after pretraining on 30.1K hours of mixed human, robot, and simulation video. The method, RynnVLA-Latent, skips native robot controls entirely: every video transition is labelled as a continuous 'latent action' instead, letting the authors fold human footage and simulation rollouts into the same supervision pipeline as actual robot demos.

The abstract frames the central move cleanly. 'RynnLAM learns continuous, 3D-grounded latent actions from visual transitions without requiring native robot controls. By combining residual future-feature prediction, camera-compensated 3D-flow supervision, and camera-motion decoupling, RynnLAM encourages representations that capture control-relevant motion rather than appearance changes alone.' Those 30.1K video hours yield 745.52M latent-action descriptors, which the policy learns to predict in chunks before fine-tuning to specific robot embodiments.

The same policy scores 84.7% on LIBERO-Plus without benchmark-specific adaptation and 51.9% average across five VLABench tracks. The paper offers no breakdown of which real-world task types succeed or fail, and the arxiv page lists fourteen authors but no institutional affiliations. It arrived during a busy stretch for the field: AI Weekly has tracked 167 robotics stories over the last 90 days, including Amap's open-source λ0 humanoid from the same week.