Moving Atoms wants to be the internet-scale data unlock for robots

Key points

Key takeaways from Moving Atoms ' launch video introducing Moving Atoms (August 2026):

The thesis: robotics is bottlenecked by data. The founders — a pair of brothers — argue teleoperation and traditional simulation won't generate enough training data for robots, and that world models are the fix, playing the role the internet's text played for LLMs.

What a world model is here. They describe it as a virtual-reality environment in which a robot's actions can be trained and evaluated, built by taking internet-scale video, making it physics-aware, and conditioning it on robot actions.

First model, Atom 1. The company's inaugural model is pitched as both physics-aware and controllable, and is claimed to top the Physics IQ benchmark, ahead of Nvidia's Cosmos 3.

Policy evaluation as the wedge. Users can submit a training policy at movingatoms.ai and have Atom 1 evaluate it across a thousand simulated environments to surface exactly where it breaks — pitched as up to 50x cheaper than testing on a real robot and saving weeks of deployment time.

Early performance claim. The founders say world-model augmentation has taken policies from 0% to 28% success in previously unseen environments.

Summary based on the company's short launch video (transcript captured); figures and benchmark claims are the company's own.