At Computex 2025, Nvidia has rolled out a major update to its robotics AI lineup. Leading the charge is Isaac GR00T N1.5, the first upgrade to its open, general-purpose foundation model built to teach humanoid robots how to think and move more like us.
GR00T N1.5 brings a smarter edge to the table—it’s now better at adapting to new environments and changing workspace layouts. It can even recognise objects based on user instructions, making interactions more intuitive. This update boosts its performance in everyday material handling and manufacturing tasks, like sorting items or stashing them in the right place, with a much higher success rate than before, as per Nvidia.
Alongside it, Nvidia introduced Isaac GR00T-Dreams—a blueprint designed to generate synthetic motion data to train robots faster and more flexibly. And to power it all, Nvidia unveiled new Blackwell systems, built to supercharge the development of next-gen humanoid robots.
Nvidia Isaac GR00T-Dreams is essentially a cheat code for training robots faster and smarter. It’s a blueprint designed to churn out massive amounts of synthetic motion data—also known as neural trajectories—that physical AI developers can tap into to teach robots new tricks, especially in dynamic, ever-changing environments.
Here’s how it works– developers first fine-tune Cosmos Predict world foundation models (WFMs) for their specific robot. Then, just by feeding in a single image, GR00T-Dreams can whip up realistic videos of the robot tackling fresh tasks in unfamiliar settings. From these, it pulls out action tokens—think of them as compact, ready-to-use nuggets of movement know-how—which are then used to train the robot to perform those tasks in real life.
The GR00T-Dreams blueprint works hand-in-hand with Isaac GR00T-Mimic, which Nvidia introduced back at GTC in March. The key difference? While GR00T-Mimic leans on the Nvidia Omniverse and Cosmos platforms to beef up existing datasets, GR00T-Dreams goes a step further, using Cosmos to generate brand-new data from scratch.
Nvidia Research used GR00T-Dreams to create the synthetic training data needed for GR00T N1.5, the next-gen upgrade to GR00T N1. Nvidia says what would’ve taken nearly three months of manual data collection was instead pulled off in a fraction of the time, thanks to Dreams’ ability to automate and accelerate the whole process.