NVIDIA's latest move on the Omniverse platform is pretty aggressive. The core is synthetic data generation—using virtual environments to train physical AI models, addressing the pain point of robots and autonomous vehicles lacking real-world data.

The key update is Cosmos Predict 2.5, which can generate multi-camera video worlds directly from a single image, video clip, or text prompt. It may not look like much, but this means developers can mass-produce realistic training data, allowing models to perform more reliably in the real world.

Some companies have already jumped on board: Skild AI uses it for robot strategy testing, Serve Robotics combines synthetic and real data to train autonomous delivery robots, and they’ve already completed over 100,000 deliveries in public spaces. Even mining companies are using it to optimize exploration systems with synthetic data.

In short, NVIDIA is using virtual data to narrow the gap between simulation and reality. If this approach scales up, it will accelerate the deployment of robots and autonomous vehicles across the board.