CLAP: Cross-Embodiment Video World Models are Zero-Shot Physical Simulators
CLAP trains video world models across human and robot videos, then grounds them for zero-shot robot simulation.
The paper says CLAP uses end-effector poses, language instructions, and latent actions to bridge mismatched action spaces across embodiments. Its curriculum first learns physical priors from unlabeled video, then ties them to robot action spaces for real-world deployment. The authors report that CLAP approaches or beats single-embodiment video models on DROID, with further gains from few-shot adaptation. They say the code and models are open-sourced. ArXiv · AI/CL/LG's note
The paper says CLAP uses end-effector poses, language instructions, and latent actions to bridge mismatched action spaces across embodiments. Its curriculum first learns physical priors from unlabeled video, then ties them to robot action spaces for real-world deployment. The authors report that CLAP approaches or beats single-embodiment video models on DROID, with further gains from few-shot adaptation. They say the code and models are open-sourced. ArXiv · AI/CL/LG's note
score 6