PhyCo: Learning Controllable Physical Priors for Generative Motion

Sriram Narayanan, Ziyu Jiang, Srinivasa G. Narasimhan, Manmohan Chandraker, [3mm] Carnegie Mellon University NEC Labs America UC San Diego [1mm] phyco-video.github.io · 2026

Read PhyCo: Learning Controllable Physical Priors for Generative Motion by Sriram Narayanan, Ziyu Jiang, Srinivasa G. Narasimhan, Manmohan Chandraker, [3mm] Carnegie Mellon University NEC Labs America UC San Diego [1mm] phyco-video.github.io (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2604-28169.txt.

Description

Narayanan's team gives video generation physics control through PhyCo, tackling drift, flat crashes and wrong matter. They build a large dataset of simulated videos varying friction, rebound, bending and force, fine-tune a diffusion model with physics-tuned control, and add vision-language reward feedback on physics queries. The result makes motion that stays true to physics and users can steer, without a sim at run time. It lifts realism on a physics benchmark and in human studies.

Catalog record

CREATORSriram Narayanan, Ziyu Jiang, Srinivasa G. Narasimhan, Manmohan Chandraker, [3mm] Carnegie Mellon University NEC Labs America UC San Diego [1mm] phyco-video.github.io
YEAR2026
COLLECTIONPAPERS
FORMATTXT
FILE SIZE49414
LICENSEhttp://creativecommons.org/licenses/by/4.0/
RIGHTSCC BY
SOURCEarXiv

Source: https://arxiv.org/abs/2604.28169

Browse the full library · About the Tower of Babel