At the Computex 2025 tech trade show, chipmaker Nvidia has made clear its intentions of forging ahead on the company’s humanoid robots program. It has now introduced an upgraded version of its humanoid robotics foundation model, Isaac GR00T N1.5, and a novel data generation framework, Isaac GR00T-Dreams. These tools are designed to dramatically reduce the time and cost of training robots for real-world tasks.
“Physical AI and robotics will bring about the next industrial revolution,” Jensen Huang, founder and CEO of Nvidia, commented on the development. “From AI brains for robots to simulated worlds to practice in or AI supercomputers for training foundation models, NVIDIA provides building blocks for every stage of the robotics development journey.”
The GR00T-Dreams system enables the generation of synthetic video content (referred to as “dreams”) by prompting models with simple inputs, such as static images and user instructions. These videos simulate the execution of specific tasks by robots in various environments, creating data that can be used to train and refine AI models. Using Cosmos Predict, GR00T-Dreams transforms these simulations into usable training data by extracting compact behavioral tokens.
These action tokens enable robots to learn new capabilities without relying on exhaustive manual demonstrations. This means that by enabling the generation of vast amounts of high-quality synthetic data (the “dreams”), GR00T-Dreams does away with the costly and time-consuming bottleneck of real-world data collection. What previously took months of manual effort can now be achieved in days or even hours.
GR00T N1.5 is built on the synthetic data from GR00T-Dreams and the latest iteration of its general-purpose humanoid AI foundation model. This update was developed using the new synthetic training framework, significantly reducing the time needed to train the model. GR00T N1.5 is built to execute complex manipulation tasks and respond to natural language commands with improved precision, flexibility, and task generalization. The chipmaker reports that the updated model performs better in dynamic and unfamiliar environments, including manufacturing settings where task variation is common. Among the first organizations to implement GR00T N1.5 are AeiRobot, Foxlink, Lightwheel, and NEURA Robotics. Their applications include everything from robotic assembly lines and automation of repetitive tasks to household robotics development.
To support these advancements, Nvidia has also expanded its computing infrastructure by introducing new Blackwell GPU-powered platforms, including the GB200 NVL72 system. This rack-scale system combines 72 Blackwell GPUs with 36 Grace CPUs, interconnected through high-bandwidth NVLink switch fabric. This hardware configuration is optimized for running large-scale AI models necessary for complex robotic reasoning and simulation. Additionally, Nvidia presented new tools to further strengthen the simulation ecosystem that underpins its robotics efforts. These include Isaac Lab 2.2, an open-source learning framework, and Cosmos Reason, a new world model focused on chain-of-thought reasoning. The latter is aimed at curating higher-quality training datasets by generating more realistic and logically coherent synthetic data.
