PhysiCalWorld: In-Context Physics Calibration for World Models

Forum topic · 小凯 · 2026-10-23

Summary

PhysiCalWorld is a research paper (arXiv:2610.30125) proposing a method to calibrate action-conditioned world models to instance-specific dynamics via in-context prompting, without fine-tuning or analytical physics modeling. Existing world models perceive scenes only through current visual frames or narrow history windows, which often fail to reveal the physical properties governing transition dynamics—such as object mass distribution, joint damping, or surface friction. Without calibrating these factors, predictions regress toward empirical priors and drift severely in novel scenes. PhysiCalWorld encodes an exploratory interaction clip collected in the test scene into a compact physics descriptor token that grounds subsequent predictions in the scene's actual physical behavior. The model trains end-to-end using only latent prediction losses, requiring no physics labels. At inference, the descriptor token is extracted once and reused across rollouts with negligible overhead. Evaluations across four environments with diverse physics variations show significantly improved rollout prediction accuracy and visual planning success rates, with the learned descriptors forming a physically structured latent space that generalizes to physics ranges unseen during training.

Paper Overview

Field: Computer Vision (CV)
Authors: Yunhao Luo, Wentao He, Yinpei Dai, Hyelin Nam, Hongze Fu, Joyce Chai, Nima Fazeli, Jeong Joon Park
Published: 2026-10-21
arXiv: 2610.30125

Abstract

Existing world models typically perceive a scene through the current visual frame or a narrow history window. However, these observations may not reveal the physical properties that determine transition dynamics, such as an object's mass distribution, joint damping, or surface friction. Without calibrating these physics factors for world models, predictions tend to regress toward an empirical prior and induce severe drift in novel test scenes.

To address this, the authors propose PhysiCalWorld, which calibrates action-conditioned world models to instance-specific dynamics via in-context prompting, without fine-tuning or analytical physics modeling.

Key Ideas

The method encodes an exploratory interaction clip collected within the test scene into a compact physics descriptor token, which grounds subsequent predictions in the scene's actual physical behavior.
Training is end-to-end using only latent prediction losses — no physics labels are required.
At inference time, the descriptor token is extracted once and can be reused across multiple rollouts, adding negligible computational overhead.

Results

Evaluated on four environments covering diverse physics variations.
PhysiCalWorld significantly improves rollout prediction accuracy and visual planning success rates.
The learned descriptors form a latent space with physical structure and generalize to physics ranges unseen during training.

---

*Auto-collected on 2026-10-23.*