Magic-W0: A Structured World-Action Foundation Model for Physical Intelligence
Magic-W0 ties robot action generation to structured predictions of 3D state, motion, and task outcome.
The paper frames robot interaction as a transition from current state to future state, mediated by action-induced 3D motion. Its architecture lets candidate actions shape world-transition predictions while those predicted representations feed back into action generation. The authors report a 27.10 average score on RoboDojo-Sim, highest among the compared world-action models. They also say fine-tuning on limited downstream data produced strong results across several real-robot tasks. HF Daily Papers' note
The paper frames robot interaction as a transition from current state to future state, mediated by action-induced 3D motion. Its architecture lets candidate actions shape world-transition predictions while those predicted representations feed back into action generation. The authors report a 27.10 average score on RoboDojo-Sim, highest among the compared world-action models. They also say fine-tuning on limited downstream data produced strong results across several real-robot tasks. HF Daily Papers' note
score 6