PhysiCalWorld calibrates world models from in-context physics clips
Original title: PhysiCalWorld: In-Context Physics Calibration for World Models
One reusable descriptor token, trained only with latent prediction loss, improves rollout accuracy and visual planning.
PhysiCalWorld calibrates action-conditioned world models to instance-specific dynamics such as mass distribution, joint damping, and friction without fine-tuning or analytical physics. An exploratory interaction clip from the test scene becomes a compact physics descriptor token that grounds later predictions. It is trained end-to-end with latent prediction loss alone and the token is extracted once for reuse. Across four environments, the authors report better rollout accuracy and visual planning success, including beyond training physics ranges.
- The descriptor is extracted once and reused across rollouts.
- No physics labels are used during training.
- Descriptors form a physically structured latent space.
- Evaluation covers four environments with varied physics.
From the original article
1 min read at arxiv.orgExisting world models typically perceive a scene through current visual frame or a narrow history window. However, these observations may not reveal the physics properties that determine transition dynamics, such as an object's mass distribution, joint damping, or surface friction.
By Yunhao Luo · arxiv.org
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.