How Digital Twins and Real-Time Simulation Are Powering the Physical AI Revolution

In this talk

1. 0:39

   Physical AI connects computing to the real world

   The presenter frames physical AI as a way to bring computing into industries that work with physical goods and environments, including agriculture, transport, construction, life sciences, and manufacturing. Robots combine sensors, AI decisions, and physical actions in a continuous loop. The talk treats autonomous vehicles, instrumented buildings and cities, industrial robots, and healthcare robots as parts of this broader field.

2. 5:58

   Three computers support robotic systems

   Building a robot requires computers for training its AI, running that AI on the robot, and simulating the world around it. Simulation can supply training examples that are costly or risky to collect in reality, and help teams test systems before deployment. The presenter points to NVIDIA DGX for training and Jetson and IGX systems for running robotics software.

3. 12:36

   Simulation helps design and operate factories

   The presenter describes factories as coordinated systems of sensors, mobile robots, robotic arms, and humanoids. Teams can model factories before construction and use simulation to check layouts, commission equipment, and validate robot behavior. Examples named include PepsiCo, Krones, SK hynix, and warehouse projects involving GXO, Siemens, and other partners. AI factories also need simulations spanning equipment, heat, networks, power grids, and substations.

4. 21:54

   Software libraries and agents broaden access to simulation

   The Omniverse team is breaking its technology into libraries that can be used within other software and by coding agents. The presenter names OVRTX for rendering and USD, OV Physics, OV Storage, and SimReady tools for preparing assets for simulation. Agents could help teams build virtual environments, operate simulations, and assess results, reducing reliance on a small pool of simulation specialists. PTC integrations with Windchill and Onshape connect product design with simulation-ready assets.

5. 31:50

   Models, synthetic data, and compute advance robotics

   The presenter describes Cosmos as a world model for understanding and predicting physical scenes, and outlines plans to combine prediction and understanding capabilities. Robotics teams can start with broad internet-scale information, then use simulated data to specialize and evaluate models in repeated training cycles. NVIDIA’s Physical AI Data Factory Blueprint supports generating, augmenting, and evaluating data; the presenter’s central claim is that more compute can help create training data without collecting every scenario in the real world.