Bringing Robots to Life with AI: The Three Computer Revolution
Yashraj Narang, head of NVIDIA's Seattle Robotics Lab, joins the NVIDIA AI Podcast to explain how robots come to life through AI. He outlines NVIDIA's three-computer strategy for robotics: DGX systems (e.g., GB200, Blackwell) for training large AI models, Omniverse and Cosmos for simulation and synthetic data generation, and Jetson AGX Thor for onboard inference. The discussion spans the lab's research across perception, planning, control, reinforcement learning, imitation learning, and vision-language-action models. Narang contrasts imitation learning (mimicking human demonstrations) with reinforcement learning (intelligent trial and error that can exceed human performance), and the modular perceive-plan-act approach with the newer end-to-end paradigm that maps sensors directly to motor commands—drawing parallels to autonomous driving's evolution toward hybrid architectures. He addresses the explosion of humanoid robotics, attributing it to the convergence of mature hardware, LLMs/VLMs, and the fact that human environments are built for human form factors, while predicting traditional and humanoid robots will coexist. On data, Narang notes robotics lacks an internet-scale corpus, making simulation essential. He details the sim-to-real gap across perception, physics, and latency, and techniques to close it (domain randomization, adaptation, invariance). NVIDIA's proposed "data pyramid" layers YouTube videos, synthetic simulator data, and real-world data. He closes with a teaser of the CoRL conference in Seoul and the lab's Neural Robot Dynamics (NeRD) project, which replaces explicit physics computations with differentiable neural networks enabling fine-tuning and speed gains on AI-optimized hardware.
Original sources
Podcast NVIDIA AI Podcast
Episode:Bringing Robots to Life with AI: The Three Computer Revolution - Ep. 274 ↗
AI Radar wrote this summary from the publisher’s notes with AI, not from a transcript. How it works.