Magic-W0: A Structured World-Action Foundation Model for Physical Intelligence
This preprint introduces Magic-W0, a robot manipulation foundation model that couples continuous action generation with explicit latent predictions of current 3D geometry, action-induced 3D motion, and future semantics. On RoboDojo-Sim it posts the highest average score among world-action models at 27.10, and on five real-robot tasks it reaches 94.6% average success after fine-tuning, beating π0.5 at 91.8%. The result matters because it shifts policies from blind imitation toward predicting how candidate actions will change the physical world, with gains in generalization and long-horizon manipulation.
Can a robot policy jointly learn structured, control-oriented world transitions—current 3D state, action-induced 3D motion, and future semantics—and use them bidirectionally with continuous action generation?
General-purpose robot policies often map observations and language to actions through implicit scene understanding. They do not explicitly constrain what state changes a candidate action would produce, so they can fail at physical reasoning, longer-horizon changes, and generalization to unseen motions or randomized scenes.
Earlier vision-language-action policies such as OpenVLA, π0.5, and Xiaomi-Robotics-1 learn action mappings from large robot datasets, acquiring dynamics only implicitly. World-action models typically predict either future images in pixel space, as in UniPi, Dreamitate, and RoboDreamer, or generic latent features, as in DINO-WM, Video Prediction Policy, and Fast-WAM. Pixel prediction wastes capacity on textures, lighting, and backgrounds unrelated to control, while generic latent prediction lacks an explicit control-oriented structure and often treats world prediction as an auxiliary objective or final conditioning feature rather than a continuously interacting partner.
Magic-W0 introduces a three-part Structured World Transition: Current State combines vision-language context with Current 3D Geometry, Transition is represented by 3D Motion, and Future State is captured by Future Semantics. It uses a Qwen3.5-2B vision-language backbone and a layer-aligned architecture with a 3D stream, a semantic stream, and a continuous flow-matching action expert. The three expert streams exchange information at six joint-attention depths across 24 layers. The model is pre-trained on roughly 2.014 million manipulation episodes from egocentric human, UMI, real-robot, and simulation data, mapped into a unified 34-dimensional state-action interface, plus about 2.61 million vision-language samples at a 9:1 manipulation-to-VLM ratio. Frozen Track4World and DINOv3 teachers provide latent supervision for geometry, motion, and semantics. Training also includes FAST discrete action and VQA losses. Action chunks have length 50, and inference uses 10 Euler integration steps.
On RoboDojo-Sim's 42 tasks, Magic-W0 attains a 27.10 average Score and 20.84% average success rate, first among compared world-action models and 9.92 points above OpenWAM-α. Its Generalization score is 28.20 with 22.01% success, topping the listed WAMs and exceeding Xiaomi-Robotics-1 by 4.66 points. On LIBERO after fine-tuning, it averages 99.1% success, with 99.4% on Spatial and 99.6% on Goal tasks. On five real-robot tasks evaluated over 500 trials, it averages 94.6% success versus π0.5 at 91.8%, a 2.8-point gain. It improves object storage from 90% to 95%, kitchen storage from 91% to 95%, pen storage by 3 points, and clothes folding by 2 points, while tying π0.5 at 100% on bottle uprighting.
The authors plan to extend the method to dexterous-hand precision manipulation, incorporate tactile and force signals, and reduce the extra computation and inference latency from multilayer world-action interaction. The paper is a preprint without peer review. Real-robot results cover five tasks in one deployment setting and mainly compare against π0.5. Simulation and real-world results are fine-tuned rather than zero-shot. World supervision depends on frozen Track4World and DINOv3 teachers, and inference speed is not benchmarked against leading VLA models.
Robot foundation model providers, OEMs building general-purpose manipulation arms, and consumer or warehouse robot makers could use this architecture for policies that reason about physical transitions. The nearest value is fine-tuned deployment in pick-place, sorting, and household tasks, likely within 1–3 years, provided inference latency is reduced and the model is validated on more embodiments and dexterous tasks. The reported open-source code and weights could accelerate adoption by integrators.
The full text is not republished here because the paper's license does not allow it. Read the original on arXiv.