Moving Atoms has beta-launched an open benchmark for robot policies across 100 randomly selected simulation tasks, the company said in a post on X. The benchmark uses frozen scenes, seeds and budgets so results can be reproduced.
The San Francisco company lets teams submit a policy for an independent assessment across all benchmark tasks and compare two policies in identical environments, according to the company's website. That gives robotics teams a common way to test policies under the same conditions.
Moving Atoms says each task has a success predicate declared before a run, and the simulator reads the final state against that predicate. Its evaluation page adds: "The policy asks to stop. The simulator decides whether it succeeded."
Moving Atoms is a two-person Y Combinator Summer 2026 company building AI world models for robot training. Brothers Shehryar Saroya, the CEO, and Ahmad Saroya, the CTO, founded the company in 2026. Both are Oxford mathematics graduates; Shehryar previously worked on TikTok's core recommendation algorithm, while Ahmad trained neural networks for GCHQ.
The company's RobotGym product is a browser-based simulation infrastructure platform for physical AI, Moving Atoms says. It first released in September 2026, with Research priced at $100 a month, Lab at $300 a month and custom Enterprise arrangements.
Moving Atoms says its core technology, Atom 1, is a world model trained on internet-scale video and fine-tuned for physics. The company says the model is intended to let robot labs train robots using GPUs rather than human-collected manual data.
