The authors trained an end-to-end model that encodes an interaction clip (a sequence of observation-action pairs) from a test scene into a single physics descriptor token.
World models calibrated to hidden physics via short interaction clips
PhysiCalWorld uses a physics descriptor token learned from video alone to condition predictions without fine-tuning.
Top University
Yunhao Luo · Wentao He · Yinpei Dai · Hyelin Nam · Hongze Fu · Joyce Chai · +2 more
University of Michigan
Research Digest··2 min read
The authors propose PhysiCalWorld, a method that calibrates action-conditioned world models to instance-specific dynamics by prompting them with a short exploratory interaction clip.
§