Nvidia enhances the development of humanoid robots with its new open model Isaac Groot N1

Nvidia has presented Isaac Great N1, The first customizable and open base model designed to contribute generalized skills and reasoning to humanoid robots, In addition to showing the open source physics engine to develop Newton robots.

Within the framework of the GTC 2025 conference, the founder and executive director of Nvidia, Jensen Huang, has underlined the Importance of the robotics industrywhile indicating that, by the end of this decade, the world will have a deficit of at least 50 million workers.

Given all this, Nvidia plans to offer a set of technologies to train, implement, simulate and test next generation robotics. Therefore, among other novelties, it has announced the availability of Isaac Groot N1, the First open and customizable base model for reasoning and humanoid skills.

As technological has detailed in a statement, Groot N1 is now available and is the first of a family of customizable models that Nvidia will previously train for subsequent use by robotics developers around the world.

Specifically, this base model presents a Dual System Architectureinspired by “the principles of human cognition.” Thus, as the company has detailed, the First system (System 1) It is a model of action of Fast thinking, which represents reflexes or human intuition and that is trained with the Nvidia Omniverse platform.

For its part, the second system (System 2) Use a Slow way of thinking for deliberate and methodical decision making. This system is able to reason about its environment and based on the instructions received to plan actions. Also, system 1 translates these plans devised by system 2 in precise and continuous robotic movements.

In this way, Groot N1 It allows to generalize common tasks, such as grabbing, moving objects to other places or transferring them from one arm to another. It also allows you to perform several steps tasks that require an extensive context and a combination of general skills. All this can be applied to use cases such as manipulation of materials in factories.

In addition, this model stands out for being customizable, so developers can train Groot N1 with real or synthetic data so that your humanoid robot carries out specific tasks. To show the capabilities of the Isaac Groot N1 model, NVIDIA has had the 1X humanoid robot, which has performed domestic tasks autonomously.

Training data and Groot N1 task evaluation scenarios are already available for download from Hugging Face and Github. Nvidia also pointed out that the Isaac Groot plane for movement generation is also available as an interactive demo in the Build.nvidia.com service and in Github.

Newton Physics Motor for Robotic Simulation

Following this line, Nvidia has also shown Newtonand open source physics engine together with Google Deepmind and Disney Research, designed for the Robotic simulationwhich allows robots to learn to handle complex tasks with greater precision.

This engine is based on the Nvidia Wrap Framework and will be compatible with simulation platforms such as Google Deepmind and Nvidia Isaac Lab. Robotic automatic learning work accelerate in more than 70 times.

Likewise, Disney Research will be one of the first companies to use Newtron for its robotic characters platform that drives next -generation entertainment robots, such as BDX droids inspired by Star Wars. However, the Newton physics engine is expected to be Available at the end of this year.

NVIDIA COSMOS Base Models

In addition to all this, Nvidia has launched the New Nvidia Cosmos Base Modelswhich present open and customizable reasoning capabilities for the Physics developmentwith what they offer to developers “an unprecedented control over the generation of the world.”

These models, driven by platforms Nvidia Omniverse and CosmosThey provide the developers of generation of massive and controllable synthetic data for robots and autonomous vehicles after training.

On the one hand, Nvidia has presented the models Cosmos Transferwhich allow the interference of structured video inputs, such as segmentation maps, deep maps, lord scanning or trajectory maps, to generate photorealist video outputs that translate into the generation of large -scale controllable synthetic data.

On the other hand, the technology has shown the models Cosmos Predictwhich were already announced within the framework of CES in January. These models generate states of virtual worlds from multimodal entries such as text, images or video. In this way, they allow multifotogram generation, predicting intermediate actions or movement trajectories when receiving start and end entry images, as the company explained.

Finally, Cosmos Reason It is other of the models presented, also with customization capabilities, has a space -timing awareness and uses the reasoning to understand video data and predict the results of the interactions in natural language.

With all this, the cosmos models are available in preview in the API catalog of NVIDIA and in VERTEX AI Model Garden de Google Cloud. Also, Cosmos Predict and Trasnfer are openly available in Hugging Gace and Github. For its part, Cosmos Reason is available in early access.

“Just as the great language models revolutionized the generative AI and agent, the Base models of the world of cosmos represent a great advance for physical AI“Jensen Huang said.