PhysiCalWorld: In-Context Physics Calibration for World Models

Key points

• Existing world models typically perceive a scene through current visual frame or a narrow history window.1
• 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.1
• Our 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.1
• We evaluate PhysiCalWorld across four environments covering diverse physics variations and show that it substantially improves rollout prediction accuracy and visual planning success rate.1

Sources (1)

• [1]PhysiCalWorld: In-Context Physics Calibration for World Models
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 21, 05:57 PM

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

  Existing world models typically perceive a scene through current visual frame or a narrow history window.

Extractive summary: sentences quoted from the sources.