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

Can a robotics world model run at 250 frames per second while NVIDIA's Cosmos runs at roughly 1 FPS? Saturn Dynamics CEO Elisa Seghetti says her company's ARC technology can do exactly that.

In this episode, we explore how world models could transform robot training, why high-quality industrial data might outperform massive datasets, and what it will actually take to get humanoid robots working reliably in factories.

We discuss:

- 250 FPS vs. 1 FPS: Saturn's claimed performance advantage over competing world models

- World models vs. simulators: How ARC differs from NVIDIA Cosmos and Isaac Sim

- Reinforcement learning: Moving robots from 70–80% success rates toward industrial reliability

- Edge AI: Why trillion-parameter models don't make economic sense for robot deployments

- Humanoid robots: Why wheeled humanoids and mobile manipulators may make more sense in factories

- The robotics supply chain: Why today's vertically integrated robot companies may eventually resemble the automotive industry

- Robot training data: Why Saturn is betting on high-quality industrial demonstrations instead of egocentric data

Seghetti also shares her predictions for the next 12–18 months, including why the robot data provider market could shrink as better models learn from fewer demonstrations.

00:00 Introduction: the race to train robots

01:00 Meet Saturn Dynamics CEO Elisa Seghetti

01:45 The different approaches to robot training

03:45 Can frontier AI models control robots?

05:20 How world models work

07:30 Making robots easier to train

09:15 Saturn ARC vs. NVIDIA Cosmos and Isaac Sim

11:10 Where humanoid robots fit in

13:45 Are robotics companies wasting money on training?

15:00 Why robotics needs a supply chain

17:20 Vertical integration and robot AI security

19:00 Safety, guardrails, and industrial deployment

21:20 The next 18 months for Saturn Dynamics

22:40 Will the robot data market shrink?

23:40 Closing thoughts

#Robotics #HumanoidRobots #PhysicalAI #WorldModels #NVIDIA #RobotLearning