NVIDIA Cosmos vs Saturn ARC: better robot training data, not more?

In this talk

1. 1:19

   World models help policies improve

   Saturn Dynamics builds world models that generate reinforcement learning environments for robotics. Sigetti sees real, synthetic, teleoperated, and egocentric data as useful, while arguing that scalable training environments can help policies move from moderate success toward the reliability needed for deployment. She says these models can improve existing robot policies without requiring a huge model to run on every robot.

2. 5:43

   Three ways to use a world model

   A world model predicts how a scene might change given an initial frame and a robot’s intended action. It can generate synthetic rollouts, including difficult actions and edge cases; help a planner compare possible actions; or predict a target state and generate actions to reach it. Saturn is currently focused on supporting other policies with generated training and testing scenarios.

3. 9:43

   ARC targets industrial tasks and speed

   Sigetti distinguishes NVIDIA Isaac Sim, which uses 3D assets for simulation, from learned world models such as Cosmos. She says Saturn’s ARC is built for industrial robotics and prioritizes physical accuracy and computational efficiency. According to Sigetti, ARC learns from fewer demonstrations and can generate at 250 frames per second, compared with roughly one frame per second for other models.

4. 15:11

   Robots will need a layered supply chain

   Sigetti compares today’s robotics industry with the automotive industry of the 1990s, when companies built entire systems in-house. She expects specialist suppliers to emerge across layers such as world models, policies, controllers, and hardware as the market matures. Koetsier raises security concerns around poisoned training data and robots operating near people; Sigetti says certification, cybersecurity testing, and guardrails will matter as systems move from business deployments toward homes.

5. 21:38

   Deployment data is the next test

   Saturn plans to expand data collection through its industrial deployments and study how more diverse, high-value examples affect model capability and generalization over the next 12 to 18 months. Sigetti says the company excludes egocentric data from its training corpus. She expects high-quality deployment data to produce new behaviors and believes the market for data providers may contract as companies demonstrate stronger results with more targeted data.