NVIDIA's Sanja Fidler on Spatial Intelligence, Cosmos, and Physical AI
In episode 269 of the NVIDIA AI Podcast, host Noah Kravitz speaks with Sanja Fidler, VP of AI Research at NVIDIA and head of the NVIDIA Spatial Intelligence Lab in Toronto, which she founded in 2018. Fidler frames spatial intelligence as "intelligence in 3D"—the 3D counterpart to LLMs and vision-language models—and positions it as the foundation for physical AI, which she argues will likely surpass generative and agentic AI in scale, echoing Jensen Huang's view that anything that moves will eventually be autonomous. Fidler traces her lab's trajectory from early content-creation work to full-stack physical AI: the four pillars of modeling (virtual world creation), physics-based simulation, 3D understanding, and action. She walks through the evolution of NVIDIA's world models—from GameGAN (Pac-Man) and DriveGAN to Video LDM in 2023—culminating in Cosmos, NVIDIA's World Foundation platform trained on real-world videos specifically targeting physics for physical AI rather than creative video. She emphasizes that Cosmos is intended to work alongside Omniverse and traditional physics solvers rather than replace them. At SIGGRAPH, NVIDIA announced 3D GRUT, adding retracing capabilities to 3D Gaussian Splatting and integrating it with Omniverse and Isaac so users can scan a room with a phone and immediately use it to train robots. Fidler identifies visual language models (VLMs) as the key reasoning breakthrough enabling robots to handle the long tail of novel scenarios, and argues the winning formula will combine traditional physics solvers, world models like Cosmos, and VLM-based reasoning. She closes with a call to the graphics community to help solve open physical-AI problems and describes the breakthrough she's waiting for: useful home robots in her lifetime.
Original sources
Podcast NVIDIA AI Podcast
Episode:AI, Spatial Intelligence, and 3D Content Creation With Sanja Fidler of NVIDIA - Ep. 269 ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.