Summary
Two Minute Papers host Károly Zsolnai-Fehér sits down with NVIDIA’s Spencer Huang to discuss the company’s latest robotics announcements and the state of physical AI. The conversation opens with three updates: Isaac Lab now integrates the Newton physics framework as NVIDIA moves away from PhysX, new DGX Station blueprints give developers a starting point for training robot policies at their desks, and Agile One SN, a reference platform built with Agile, ships with Jetson Thor integration, an optimized software stack, and the GR00T model ready out of the box.
Huang explains why physical AI is so much harder than large language models. Text, images, and audio exist at enormous scale, but there is no comparable dataset capturing the physics of how a body and the world push back on each other, including contact, deformation, and bidirectional coupling. The discussion also covers whole-body control and the layered model idea of fast reactive “system zero” controllers beneath the slower reasoning of vision-language-action models, and why a robot cannot simply wait on a large LLM to decide how to stay upright.
The interview also looks at neural simulation and world models, including how NVIDIA Cosmos is being blended with conventional simulation in Omniverse, Isaac Sim, and Isaac Lab. It is a useful overview for anyone following robotics, simulation, and the path toward general-purpose humanoid robots.
📺 Source: Two Minute Papers · Published November 06, 2026
🏷️ Format: Interview
