OneWM Couples World Model and Policy, Tops RoboCasa365 at 36.68%

A new arXiv paper asks whether learning to predict video and learning to produce actions can help each other inside the same causal model. The authors say yes on both sides, and lean hardest on one simulation benchmark to make the point.

On RoboCasa365, a mobile-manipulation testbed, their framework OneWM reaches 36.68% overall success. The same table lists NVIDIA's GR00T N1.6 at 21.90%, GigaWorld-Policy at 20.70%, π0.5 at 16.90% and π0 at 14.80%. An action-masked variant of OneWM's own architecture (WAM mask) sits at 28.32%. OneWM is described as "a simple framework for unified causal training that treats actions as both conditions for world prediction and targets for policy learning within a shared video–action architecture" — a mixture-of-transformers backbone with the video expert initialized from Wan2.2-TI2V-5B.

The LIBERO-Plus results carry the generalization claim. OneWM lands at 59.27% overall, versus 53.37% and 52.50% for its two internal ablations, 49.70% for Fast-WAM, 53.60% for π0 and 25.00% for WorldVLA. The paper reports it "outperforms both WAM variants on in-distribution LIBERO and out-of-distribution LIBERO-Plus," with the biggest gains under sensor noise, camera and layout perturbations.

The reciprocal direction is the quieter half of the paper. Using the identical backbone, adding the action-prediction objective improves PSNR, SSIM, LPIPS and FVD across all three benchmarks: "action learning improves multi-view world-prediction fidelity under the same architecture."

A late section called "Endless Dream" chains the two inference modes without extra training. On LIBERO the model runs 16 rounds, each generating 32 actions and eight future frames, "yielding 512 actions over a modeled horizon of 25.6 seconds." The paper does not report real-robot results for that loop. It arrives in a dense week for robotic world-action modeling that also saw VLA-Δ on state deltas and Nvidia and KAIST's ACCRUE training speedups.