London startup Saturn Dynamics launches Arc for robots

Saturn Dynamics has launched Arc, a world model for training and testing robots in interactive simulations. The London startup says the system outperformed comparable NVIDIA models in its benchmark tests.

Arc is designed to help robotics teams test how machines respond to actions in simulated environments before deployment on physical hardware. The approach is intended to reduce costly, time-consuming training on real machines, where errors can damage equipment or create safety risks.

The company was founded in London by a team with backgrounds in theoretical physics, reinforcement learning and startup building. It says Arc was built to model how the physical world behaves, rather than relying mainly on video or image-based training data.

That distinction matters for developers trying to move AI systems from controlled demonstrations into factories, warehouses and other workplaces. Robots can be trained to recognise objects from visual data, but they also need to predict what happens when they push, grasp, drop or otherwise interact with them.

Saturn Dynamics is entering a market where several groups are trying to build so-called world models for physical AI. Better-known efforts include NVIDIA's Cosmos platform and Google DeepMind's Genie 3, both of which focus on generating environments in which machines can learn through interaction.

Benchmark results

In benchmark tests cited by the company, Arc achieved 73% greater trajectory accuracy than an NVIDIA Cosmos model. In other words, Arc's predicted trajectories were closer to the reference paths used in the tests.

In a separate generation benchmark using NVIDIA Jetson Thor hardware, Arc generated one second of data in 0.6 seconds, according to the company. That compared with 23 seconds for NVIDIA Cosmos Nano and 9 seconds for NVIDIA Cosmos Edge, using NVIDIA's publicly released code.

For robotics developers, generation speed affects how many scenarios can be tested in a given period. Faster simulation allows teams to run more training cycles and examine more edge cases before a machine reaches a real-world setting.

Saturn Dynamics argues that the sector has often treated realism and efficiency as a trade-off, with more detailed physical simulation requiring far more compute and time. It says Arc was developed to challenge that assumption by embedding mathematical symmetries and physical principles into the model architecture and training approach.

Physical AI push

The market for physical AI is drawing growing attention as more machines move into industrial and service settings. Around five million robots now operate in factories worldwide, according to figures cited by Saturn Dynamics, while interest in humanoid and autonomous systems has widened across manufacturers and logistics groups.

Yet deployment remains limited. McKinsey said in a 2025 report that about 50 companies had declared ambitions in humanoid robotics, but fewer than ten had moved to scaled pilots or pre-commercial deployments.

IBM has also identified several barriers to training physical AI systems, including the cost of gathering real-world data, the difficulty of reproducing physical behaviour accurately, the need for real-time response, and the fact that mistakes in live environments can have serious consequences.

Elisa Seghetti, Chief Executive Officer and Co-founder of Saturn Dynamics, said: "Chat models dominated the first wave of AI, but the next frontier will extend far beyond our screens, with intelligent robots working alongside us in factories, hospitals and homes. These systems need to understand much more than language. They need to understand how objects move, how surfaces behave and how the physical world responds to their actions. A convincing demonstration is only a starting point. Robots need to cope when things don't go to plan. With Arc, we're building a way for robotics teams to simulate these situations and learn from them before deploying machines in the real world. By building an understanding of physics into the model from the outset, we aim to make this process faster and less reliant on enormous amounts of data and computing power, helping bring the potential of physical AI closer to reality."

Funding and team

Saturn Dynamics was founded in 2025 and has raised USD $4.3 million, according to the company. Its investors include Twin Path Ventures, Vento, Sequel, Portfolio Ventures, Look AI Ventures and Masia, alongside angel investors working at groups including DHL, Datadog, McKinsey & Co, Maersk, Perk and Pirelli.

Seghetti, 26, previously founded an AI company at 21. The senior team also includes Co-founder and Chief Technology Officer Simone Totaro, who studied for a PhD in reinforcement learning, and Co-founder and Chief Scientific Officer Cesc Cunillera, who holds a PhD in theoretical and mathematical physics.

Saturn Dynamics is targeting robotics teams, manufacturers and system integrators seeking to test autonomous systems before putting them to work in practical settings.