Saturn Dynamics, a London-based startup founded in 2025, has built Arc, a physics-based world model designed to let robotics developers test actions in simulation before risking them on physical hardware. The company says Arc delivers 73% greater trajectory accuracy than Nvidia's Cosmos and generates one second of simulation data in 0.6 seconds on Nvidia Jetson Thor hardware — compared with 23 seconds for Cosmos Nano and 9 seconds for Cosmos Edge, according to Robotics and Automation News. The startup has raised $4.3 million from Twin Path Ventures, Vento, Sequel and Portfolio Ventures.
Why Speed and Physics Accuracy Both Matter for World Models
World models are the component robotics companies use to predict what happens next when a robot grips, pushes or drops something, without running that experiment on a real, breakable arm. Most existing models, including Nvidia's Cosmos family, lean on large-scale video prediction trained to look plausible. Saturn Dynamics' pitch is different: Arc bakes mathematical symmetries and physical principles directly into the model's architecture, rather than hoping a large enough video dataset teaches physics implicitly.
That distinction shows up in the benchmark Saturn Dynamics pushes hardest — generation speed on embedded hardware. A model that takes 23 seconds to simulate one second of action, as Cosmos Nano reportedly does on Jetson Thor, is unusable for testing a robot's next move in anything close to real time. Arc's 0.6-second figure, if it holds up outside Saturn Dynamics' own benchmarks, would make on-device simulation viable for robots evaluating options mid-task rather than offline in a data center. Co-founder and CEO Elisa Seghetti put the stakes plainly: "A convincing demonstration is only a starting point. Robots need to cope when things don't go to plan."
A Crowded, Underfunded Corner of Physical AI
Saturn Dynamics is a minnow next to the capital being thrown at world models elsewhere in the industry — AI research lab Emulate alone raised a $700 million seed round for the same category, and Nvidia has made Cosmos a flagship platform backed by its hardware ecosystem. Four million dollars buys a small team and a narrow technical bet, not a sales force or a hardware partnership roster.
What Saturn Dynamics has that bigger rivals don't is a specific, falsifiable claim: physics-grounded architecture beats video-scale pretraining on both accuracy and inference speed, at least on the benchmarks it has published. If independent robotics integrators can reproduce those numbers on their own tasks, Saturn Dynamics becomes an attractive acquisition target for a larger robotics or chip company looking to shortcut world-model development rather than out-spend Nvidia. If the numbers don't generalize beyond Saturn's own test suite, the company joins a long list of physical AI startups whose benchmark slide outran its product.
