Magic-W0: A Structured World-Action Foundation Model for Physical Intelligence
Magic-W0 jointly models structured 3D world transitions and robot actions, leading compared world-action models.
Magic-W0 is a world-action foundation model that jointly predicts structured physical-state transitions and continuous robot actions. Its Structured World Transition combines vision-language context and current 3D geometry, action-induced 3D motion, and future task semantics, with layer-aligned coupling between action hypotheses and world prediction. It is pretrained on large-scale egocentric human manipulation, UMI, real-robot, and simulation data with latent supervision from pretrained visual models. On RoboDojo-Sim it scores 27.10, the highest among compared world-action models, and fine-tunes well on real-robot tasks with limited data.
- Magic-W0 models Current State, 3D Motion Transition, and Future Semantics together.
- Action hypotheses condition world prediction, and world features inform action generation.
- Pretraining uses egocentric human, UMI, real-robot, and simulation data.
- It scores 27.10 on RoboDojo-Sim, highest among compared world-action models.
From the original article
1 min read at huggingface.co
World-action models (WAMs) augment robot policies with action-conditioned environment dynamics, yet existing approaches largely rely on future observation reconstruction or generic latent prediction and lack structured, control-oriented world representations tightly coupled with action generation.
By Xuhua Chen, Zhenhan Yin, Yuan Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.