Nvidia has launched a "brain" capable of reasoning for robots, and the upgraded Cosmos world model is here.

Huang (Jensen Huang) is really serious about his optimism for robots!

Right at the ongoing SIGGRAPH (Computer Graphics) Conference, NVIDIA introduced a fully upgraded Cosmos world model for robots.

Cosmos is mainly used to generate synthetic data that adheres to the physical laws of the real world. Since its release, it has been adopted by well - known robotics and autonomous driving companies such as Figure, Agility Robotics, and General Motors.

This upgrade focuses on two major aspects: planning ability and generation speed.

• Launch of Cosmos Reason: A vision - language model with 7 billion parameters and inference capabilities, which can assist robots in task planning.
• Introduction of Cosmos Transfer - 2 & its Lite version: It accelerates the conversion of virtual scenarios into training data and offers a faster, lightweight version.

In addition to software updates, NVIDIA has almost completely upgraded the supporting technologies and hardware for robot development, including a new Omniverse library, RTX PRO Blackwell servers, and the DGX Cloud platform.

Sanja Fidler, vice - president of NVIDIA's artificial intelligence research, said:

AI is enhancing our simulation capabilities, and simulation capabilities are driving the development of AI systems. There is a real and powerful combination between these two fields, which few can achieve.

Now, NVIDIA is leveraging a comprehensive set of robot development infrastructure to make a full - force push towards this new frontier of integration.

Both software and hardware have been updated

Let's first take a look at what specific updates there are.

On the software side, Cosmos Reason, which is only 7B in size, is particularly noteworthy.

Officially, the model not only has a solid understanding of physics but also can use common sense for multi - step reasoning, serving as the "brain" of robots.

Through memory and reasoning, it can act as a planning model to infer the next possible steps of physical agents.

Meanwhile, NVIDIA has also launched Cosmos Transfer - 2 for accelerating synthetic data generation and its lighter, streamlined version, significantly reducing the conversion time from virtual scenarios or spatial control information to training data.

The streamlined version simplifies the 70 - step distillation process to one step, enabling it to run at a higher speed on RTX PRO servers.

To support these models, NVIDIA has also released a more powerful new Omniverse library and SDK, allowing robot developers to more conveniently perform 3D reconstruction, simulation, and cross - platform data interoperability. Specifically, it includes the following:

1. Data compatibility between MuJoCo and OpenUSD, two robot simulation formats has been achieved, enabling a large number of developers to more easily conduct robot simulations across different platforms.

2. Omniverse RTX ray - tracing and 3D Gaussian rendering technologies have been introduced, allowing developers to capture, reconstruct, and simulate the real world in 3D using sensor data.

3. The robot simulation and learning tools Isaac Sim 5.0 and Isaac Lab 2.2 have been open - sourced on GitHub. Combining new rendering technologies and a unified data format, they help robot developers better bridge the gap between virtual simulation and the real environment.

In short, Cosmos and Omniverse are a golden pair (the former is responsible for building realistic and efficient 3D simulation and digital twin environments, while the latter focuses on providing powerful physical AI models and intelligent reasoning capabilities). When combined, they can generate more accurate and large - scale data for robot training.

After dealing with data, there is the issue of hardware.

In response, NVIDIA has specifically launched the RTX PRO Blackwell server, which covers workloads such as training, synthetic data generation, robot learning, and simulation.

Moreover, with DGX Cloud, developers can efficiently run Omniverse and Cosmos - related applications via the cloud anytime, anywhere, easily achieving large - scale simulation and training tasks, greatly reducing hardware thresholds and operation and maintenance costs. (Currently, DGX Cloud is available on the Microsoft Cloud Marketplace.)

NVIDIA is fully betting on the robotics field

From the above actions, it is clear that NVIDIA is vigorously deploying in the field of robot R & D.

As for the underlying considerations, NVIDIA detailed them in a recent blog. In summary:

The integration of computer graphics and AI is fundamentally changing the robotics field.

Indeed, NVIDIA, which started with graphics, has chosen robotics as its next focus due to the observed trend of the integration of graphics and AI and its significant impact on the robotics field.

For example, traditional robot training relies on real - world data, which is costly and limited to obtain. Computer graphics technology can create realistic virtual environments and, combined with AI, achieve physically accurate simulations.

In this way, the threshold and cost for robots to learn various skills are reduced, making it more likely to accelerate the implementation of robot applications.

According to Ming - Yu Liu, vice - president of NVIDIA's research, NVIDIA's goal is to build a complete, realistic, and scalable "virtual parallel universe" where robots can safely experiment repeatedly and continuously evolve.

Physical AI requires a realistic virtual environment where robots can repeatedly experiment and learn in this safe parallel world.

To build such a virtual world, we must combine real - time rendering, computer vision, physical motion simulation, as well as 2D and 3D generative AI and AI reasoning technologies. These are the core areas that NVIDIA's research team has been continuously focusing on and refining over the past two decades.

Obviously, once this "virtual world for robots" is successfully built, NVIDIA can empower more developers and enterprises with this infrastructure, helping them to develop various robots and AI applications more quickly.

In this way, NVIDIA can not only further consolidate its leading position in the graphics field but also continue to play the role of a "shovel seller" in the era of embodied intelligence.

Well, what a familiar "routine" (doge)~

Moreover, from Huang's recent series of public actions, it seems that he has a particular preference for Chinese robots.

Not only has he praised them on various occasions, but recently, just as he did with OpenAI in the early days (providing supercomputers to OpenAI), he has sent the first batch of Jetson Thor chips to the domestic player Galaxy Universal.

In addition, NVIDIA is also reaching in - depth cooperation with companies such as Alibaba Cloud, Beijing Humanoid Robot Innovation Center, Fourier Intelligence, Accelerated Evolution, Ubtech Robotics, and Zhipu AI. Alibaba Cloud has even announced that NVIDIA's full - stack robot and physical AI components are available for use via its cloud services.

It has to be said that NVIDIA's business of selling "shovels" for robots in China is gradually taking off.

Open - source address of Cosmos: https://github.com/nvidia - cosmos

Reference links:

[1]https://techcrunch.com/2025/08/11/nvidia - unveils - new - cosmos - world - models - other - infra - for - physical - applications - of - ai/

[2]https://investor.nvidia.com/news/press - release - details/2025/NVIDIA - Opens - Portals - to - World - of - Robotics - With - New - Omniverse - Libraries - Cosmos - Physical - AI - Models - and - AI - Computing - Infrastructure/default.aspx

This article is from the WeChat public account "Quantum Bit". Author: Yishui. Republished by 36Kr with permission.