Inference by Turing Post Spencer Huang: NVIDIA’s Big Plan for Physical AI: Simulation, World Models, and the 3 Computers
4 snips
Dec 4, 2025 In a captivating discussion, Spencer Huang, NVIDIA’s product lead for robotics software, dives deep into the future of robotics and simulation. He outlines NVIDIA's innovative three-computer vision—training, simulation, and deployment. Spencer emphasizes the critical role of simulation in ensuring safety and speed in robot deployment. He also explores the fascinating contrast between conventional and neural simulators, tackling data bottlenecks in robotics while advocating for an open-source ecosystem. It's a thoughtful look at how robots learn and interact with the real world!
AI Snips
Chapters
Books
Transcript
Episode notes
Three Computers Power Physical AI
- Robotics needs three distinct computing tiers: training (DGX), simulation (OVX/Omniverse), and on-robot deployment compute.
- Using all three enables fast training, high-fidelity validation, and safe, redundant runtime behavior for robots interacting with people.
Simulation As The Robot Classroom
- Simulation accelerates robot learning by running faster-than-real-time training and by providing high-fidelity validation before deployment.
- Domain randomization across many simulated universes yields robustness so robots generalize to novel real-world conditions.
Robots Learn Practical Physics
- Robots learn physics via randomized training that varies physical parameters like friction and mass.
- This equips policies to adapt to different surfaces and object behaviors, creating practical physical understanding.
