CLAP: Cross-Embodiment Video World Models are Zero-Shot Physical Simulators

Authors/institution: Kechen Liu and Ola Shorinwa; Princeton University. Status: arXiv preprint, 27 August 2026. Problem: robot world models are usually tied to one embodiment. Contribution: combines end-effector, language and latent action representations to learn across human and robot videos; code and models are open. Limitations: zero-shot and benchmark claims are author-reported and deployment safety is not established. Paper.