OneWM: Unified Causal Training for World and Action Models

OneWM treats actions as both conditions for world prediction and targets for policy learning within a shared video-action architecture. Experiments on LIBERO and RoboCasa365 show that learning forward dynamics improves policy performance.