Moving Atoms develops world models using vast amounts of internet-scale video data to understand and predict physical interactions. These models are designed to respond to robot actions, enabling more accurate simulations of real-world dynamics. Their initial model, Atom 1, has achieved top performance on the Physics IQ benchmark, demonstrating strong capabilities in reasoning about physical systems. The approach leverages the inherent structure of visual data to learn fundamental laws of motion and object behavior without explicit programming.
The company views online video as a rich source of latent physical knowledge, similar to how text data unlocked large language models. By extracting this information, Moving Atoms aims to create foundational models that empower advanced robotics. This strategy positions world modeling as a critical step toward building intelligent robots capable of complex, adaptive tasks. The vision is to unlock new levels of robotic autonomy through scalable, data-driven learning from the global video ecosystem.