Physical AI startup Saturn Dynamics is developing a world model designed to help robots learn how the physical world behaves through interactive simulations, potentially reducing the time and computing resources required for training.
Founded in 2025 and based in London, UK, the company has developed Arc, a model that enables robotics developers to test actions and evaluate their consequences in virtual environments before attempting them on physical machines.
Unlike approaches that rely primarily on video demonstrations or conventional physics simulations, Saturn Dynamics says Arc incorporates mathematical symmetries and physical principles into its architecture.
The approach is intended to help machines understand how objects behave when manipulated, rather than simply recognizing their appearance or movement.
For example, developers could use Arc to test different ways of opening a drawer or handling objects on a production line, including situations where objects slip or behave unexpectedly.
Faster simulation and improved accuracy
Saturn Dynamics reports that Arc achieved 73 percent greater trajectory accuracy than an Nvidia Cosmos model in benchmark testing.
In separate tests using Nvidia Jetson Thor hardware, Arc generated one second of simulation data in 0.6 seconds, compared with 23 seconds for Cosmos Nano and nine seconds for Cosmos Edge.
The company says these results could allow developers to evaluate more scenarios while reducing computing requirements.
Elisa Seghetti, co-founder and CEO of Saturn Dynamics, said: “A convincing demonstration is only a starting point. Robots need to cope when things don’t go to plan.”
The company is targeting robotics manufacturers, developers and systems integrators seeking to improve the reliability of autonomous machines.
Saturn Dynamics has raised $4.3 million as of October 2026, with investors including Twin Path Ventures, Vento, Sequel and Portfolio Ventures.
Its longer-term objective is to help narrow the gap between simulated training and real-world performance, one of the principal challenges facing the commercialization of physical AI.
