Nvidia charges ahead with humanoid robotics aided by the cloud

Nvidia said it is racing ahead with humanoid robotics technology, providing a custom foundation model for humanoid reasoning, a blueprint for generating synthetic motion data, and more Blackwell systems to accelerate humanoid robot development.

At the Computex 2025 trade show in Taiwan, Nvidia unveiled Isaac GR00T N1.5, the first update to Nvidia’s open, generalized, fully customizable foundation model for humanoid reasoning and skills; Nvidia Isaac GR00T-Dreams, a blueprint for generating synthetic motion data; and Nvidia Blackwell systems to accelerate humanoid robot development.

Humanoid and robotics developers Agility Robotics,, Boston Dynamics, Fourier, Foxlink, Galbot, Mentee Robotics, NEURA Robotics, General Robotics, Skild AI and XPENG Robotics are adopting Nvidia Isaac platform technologies to advance humanoid robot development and deployment.

“Physical AI and robotics will bring about the next industrial revolution,” said Jensen Huang, CEO of Nvidia, in a statement. “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.”

New Isaac GR00T Data Generation Blueprint Closes the Data Gap

Showcased in Huang’s Computex keynote address, Nvidia Isaac GR00T-Dreams is a blueprint that helps generate vast amounts of synthetic motion data — aka neural trajectories — that physical AI developers can use to teach robots new behaviors, including how to adapt to changing environments.

Developers can first post-train Cosmos Predict world foundation models (WFMs) for their robot. Then, using a single image as the input, GR00T-Dreams generates videos of the robot performing new tasks in new environments. The blueprint then extracts action tokens — compressed, digestible pieces of data — that are used to teach robots how to perform these new tasks.

The GR00T-Dreams blueprint complements the Isaac GR00T-Mimic blueprint, which was released at the Nvidia GTC conference in March. While GR00T-Mimic uses the Nvidia Omniverse and Nvidia Cosmos platforms to augment existing data, GR00T-Dreams uses Cosmos to generate entirely new data.

Jim Fan, director of AI and distinguished scientist at Nvidia, said in a press briefing, “Nvidia has a very strong robotic strategy, and it is centered around what Jensen calls the three computer problem.”

He noted the firm has the OVX computer that is meant to do simulation and graphic simulation physics engine, and it is used to synthesize and generate data. And this data is consumed by the DGX computer, which are used to train foundation models. And then it is deployed to the HX computer, which is the runtime on the edge for platforms like humanoid robots.

Gr00t is the lifecycle of physical AI and robot-based workflows, Fan said.

“It is an instantiation of the three-computer problem,” he said.

He pointed to two advances in Project Gr00t, Gr00t Dreams and Gr00t N1.5. (He said he was quite proud of those names).

For Gr00t Dreams, Fan said it is a model that can generate videos to train robots. He showed a bunch of videos and said all of the videos were generated by Nvidia Cosmos.

“We found a way to apply advanced video generation models like Cosmos to help humanoid robotics. So on a high level, how this method works is we first fine tune Cosmos on robot videos from our lab so that this video model is now customized to the robots at our lab, and then we can use this fine tuned model to generate, in principle, infinite number of dream videos by prompting the model in different ways,” Fan said. “And now that becomes synthetic data to augment our real robot data sets. As many of you might know, collecting data on the real robot is very time consuming and costly, because you are fundamentally limited by 24 hours per robot per day, right? It’s a physical system, but with Gr00t Dreams, this new workflow, this new set of algorithms, now we’re able to break this fundamental physical limit and then multiply data at unprecedented scale next.”

The result is that robots will be able to pick up objects correctly. You can tell it to pick up a cucumber or pour some orange juice or open a laptop. The robot has never been trained on these particular actions, Fan said, but because it has been trained with video models, the robot is able to “understand the physics and the meaning of these verbs.”

And so it learns how to perform the actions.

New Isaac GR00T Models Advance Humanoid Robot Development

Nvidia Research used the GR00T-Dreams blueprint to generate synthetic training data to develop GR00T N1.5 — an update to GR00T N1 — in just 36 hours, compared with what would have taken nearly three months without the blueprint.

GR00T N1.5 can better adapt to new environments and workspace congurations, as well as recognize objects through user instructions. This update signicantly improves the model’s success rate for common material handling and manufacturing tasks like sorting or putting away objects. GR00T N1.5 can be deployed on Jetson Thor, launching later this year.

The Gr00t N1.5 foundation model is incorporating Gr00t Dreams as part of the synthetic data generation pipeline. Nvidia has upgraded the visual language backbone, so Gr00t N1.5 will have better adaptability and the language instruction compliance, Fan said.

Gr00t N1.5 it will be announced at Computex and then be released open source by June 9. As for Gr00t Dreams, Nvidia is still working on the timeline. The hope is to open source as much as possible, Fan said.

Early adopters of GR00T N include AeiRobot, Foxlink, Lightwheel and NEURA Robotics. AeiRobot employs the model to enable ALICE4 to understand natural language instructions and execute complex pick-and-place workows in industrial settings. Foxlink Group is using it to improve industrial robot manipulator exibility and eciency, while Lightwheel is harnessing it to validate synthetic data for faster humanoid robot deployment in factories. NEURA Robotics is evaluating the model to accelerate its development of household automation.

New Robot Simulation and Data Generation Frameworks Accelerate Training Pipelines

Developing highly skilled humanoid robots requires a massive amount of diverse data, which is costly to capture and process. Robots need to be tested in the physical world, which can present costs and risk.

To help close the data and testing gap, Nvidia unveiled the following simulation technologies:● Nvidia Cosmos Reason, a new WFM that uses chain of thoughts reasoning to help curate accurate, higher-quality synthetic data for physical AI model training, is now available on Hugging Face● Cosmos Predict 2, used in GR00T dreams, is also coming soon to Hugging Face with performance enhancements for high-quality world generation and reduced hallucination.● Nvidia Isaac GR00T-Mimic, a blueprint for generating exponentially large quantities of synthetic motion trajectories for robot manipulation, using just a few human demonstrations.● Open-Source Physical AI Dataset, which now includes 24,000 high-quality humanoid robot motion trajectories used to develop GR00T N models.● Nvidia Isaac Sim 5.0, a simulation and synthetic data generation framework, now openly available on GitHub. ● Nvidia Isaac Lab 2.2, an open-source robot learning framework, which will include new evaluation environments to help developers test GR00T N models.

Foxconn and Foxlink are using the GR00T-Mimic blueprint for synthetic motion manipulation generation to accelerate their robotics training pipelines. Agility Robotics, Boston Dynamics, Fourier, Mentee Robotics, NEURA Robotics and XPENG Robotics are simulating and training their humanoid robots using Nvidia Isaac Sim and Isaac Lab. Skild AI is using the simulation frameworks to develop general robotintelligence, and General Robotics is integrating them into its robot intelligence platform.

Universal Blackwell Systems for Robot Developers

Global systems manufacturers are building Nvidia RTX PRO 6000 workstations and servers, oering a single architecture to easily run every robot development workload across training, synthetic data generation, robot learning and simulation.

Cisco, Dell Technologies, Hewlett Packard Enterprise, Lenovo and Supermicro announced Nvidia RTX PRO 6000 Blackwell-powered servers, and Dell Technologies and Lenovo announced Nvidia RTX PRO 6000 Blackwell-powered workstations.

When more compute is required to run large-scale training or data generation workloads, developers can tap into Nvidia Blackwell systems like GB200 NVL72 — available with Nvidia DGX Cloud on leading cloud providers and Nvidia Cloud Partners — to achieve up to 18x greater performance for data processing.