PhysiCalWorld: In-Context Physics Calibration for World Models
An exploratory interaction clip lets a world model adapt to a scene’s physics without fine-tuning or explicit physics modeling.
PhysiCalWorld compresses the clip into a physics descriptor token that guides later predictions. The token is extracted once per test scene and reused across rollouts with negligible compute overhead. Training uses only latent prediction loss, without physics labels. Across four environments, the authors report substantially improved rollout accuracy and visual planning success, alongside generalization beyond the physics ranges used in training.
ArXiv · AI/CL/LG's note
PhysiCalWorld compresses the clip into a physics descriptor token that guides later predictions. The token is extracted once per test scene and reused across rollouts with negligible compute overhead. Training uses only latent prediction loss, without physics labels. Across four environments, the authors report substantially improved rollout accuracy and visual planning success, alongside generalization beyond the physics ranges used in training.
ArXiv · AI/CL/LG's note
score 5