# Nvidia CEO Shows Three-Computer Framework for Robot AI in Berlin By Simon Yoon Canonical URL: https://www.tokenpost.com/news/technology/37259 Published: 2026-10-22T01:05:01.000Z Updated: 2026-10-22T01:05:01.000Z Section: Technology > Four Microducks illustrated a workflow linking RTX Spark, cloud training and simulation before deployment to a physical robot. Nvidia CEO Jensen Huang used a Berlin keynote on Wednesday, Oct. 21, to demonstrate a robotics workflow linking model training, simulation and computing on a physical machine, showing how the company’s framework connects three stages. Four Microducks joined Huang onstage. He described them as created on RTX Spark, trained in the cloud and taught in simulation. “This demo runs on three computers,” Huang said. A software agent helped create a simulated course, and a learned policy returned from cloud training for evaluation before physical deployment. The sequence gave a live example of a framework Nvidia first outlined Oct. 23, 2024, and updated Aug. 8, 2025. The concept predates the Berlin event; the Microduck project showed how its stages can connect in a demonstration. The framework separates robotics work across training, simulation and on-robot operation. DGX systems handle training; Omniverse and Cosmos provide simulation on RTX PRO servers; Jetson AGX Thor handles inference on a robot. Inference is the stage when a trained model runs to process information and produce actions. The stages place computing in a data center, a virtual environment and on the physical machine. Training, virtual testing and execution on the device are distinct tasks that a developer’s workflow connects. Physical AI refers to systems that perceive, reason and act in the physical world. A simulated setting lets developers test behavior before deployment, while on-robot computing runs the system in its operating environment. The three-part approach links development work to the robot’s eventual use. For Nvidia, the architecture offers a potential role at multiple points in robotics development, from data-center training to simulation and computing on deployed machines. That is a business opportunity implied by the design, not a measure of demand. The keynote did not quantify adoption or revenue from physical AI. The demonstration showed how the workflow fits together; the three-computer concept itself was not a new launch. Huang’s presentation came a day after his remarks about [companies pairing factories with digital twins]() at a Berlin summit. ## Links in this article - [companies pairing factories with digital twins](https://www.tokenpost.com/news/technology/35820)