Spencer Huang Discusses NVIDIA's Physical AI Stack for Robotics

In this talk

1. 1:24

   Physical AI needs its own infrastructure

   Huang describes a progression from autonomous vehicles to humanoid robotics and says the required technologies are starting to take shape. Physical AI needs infrastructure that can simulate and test policies, then run them on real hardware before deployment. He says this hardware-in-the-loop process differs from the more homogeneous compute used for large language models.

2. 3:28

   Compute and demonstrations are advancing robot learning

   Huang attributes recent progress in robot policies partly to greater access to compute and software optimized for it. He discusses flow matching for generating movement paths and DeepMimic, which uses human demonstrations to help train actions such as running, breakdancing, and parkour. For manipulation, he says imitation learning can be brittle while reinforcement learning can require extensive tuning, and combining them in simulation can improve robustness.

3. 10:16

   Open platforms can build a shared robotics baseline

   Huang says NVIDIA aims to open-source robotics platforms and models so teams can adapt them to varied domains, from manufacturing and warehousing to underwater, aerial, and space applications. He describes Cosmos and GR00T as base models that users can post-train for specific tasks, and says data quality and commercial licensing affect where each model can be used. A broader shared dataset could improve the generalist baseline, while valuable task-specific expertise still requires more data.

4. 16:35

   Companies need focused applications and skill paths

   Huang groups robotics companies into those building hardware, brains, or both. Hardware teams need to settle on suitable robot morphology and sensors for their deployment setting, while brain-focused teams need a practical industry beachhead and compatible robot embodiments. He recommends choosing applications where basic skills can build into longer, more complex tasks and can be evaluated.

5. 20:52

   Care could be an early home and hospital use

   Huang sees assistance for older adults as an early home use, including carrying medicines, food, and blankets and responding to physical or emotional states. He also points to hospitals, where robots could take on some tasks and help overworked nurses focus on patient care. He says these uses do not require superhuman dexterity.