GTC was mostly updates, and one of them is a humanoid reference
Milos Stanic, senior solution architect at NVIDIA, opened the October physical AI special with a short GTC update. His background is simulation and CAE, including automotive, and since joining NVIDIA the work has shifted toward agents. The point of the ten minutes was context: what changed for physical AI, and where partners pick it up in practice.
The largest announcement he called out was a reference platform for humanoids, shipping with GR00T and Jetson Thor, with distribution aimed at early 2027. Around it: a new Isaac Lab and Isaac OS, Hyperion as the end-to-end autonomous-vehicle platform aimed at a higher autonomy stage, and Omniverse moving off a monolith toward libraries, kept backward compatible. NVIDIA also opened a large set of physical-AI assets, including more than 2,000 SimReady models, plus updates to Cosmos. Safety showed up as Halos for robots and vehicles, and OpenShell for agents.
Three computers, then agents across the pipeline
The frame he wanted the room to keep is the three-computer problem. Physical AI now needs one computer to train models, one to simulate, and one to deploy. On the train side he pointed at Vera Rubin in production, plus smaller systems such as DGX Spark and RTX Spark that sit closer to simulation. Simulation is the Omniverse libraries, the new Isaac Sim and Isaac Lab, and the SimReady assets. Deploy is the robot brain: Jetson Thor, Isaac OS, and Hyperion for vehicles.
What is new on top of that stack is orchestration. Stanic’s line was that where there is agentic AI there is an opportunity for physical AI, and the reverse is now obvious: the two are meeting. Agents, with skills, take the pipeline from planning through deployment. He pointed at more than 400 skills on build.nvidia.com, built so agents can help construct physical-AI assets. The GTC keynote example he sent people to watch is a single prompt that picks up the scene, simulates robot behaviour, trains, evaluates, and deploys onto a small robot.
