Nvidia Unveils the Future of Physical AI at SIGGRAPH 2025 Conference

At SIGGRAPH 2025, Nvidia is unveiling a series of innovations, including its compact Blackwell GPUs, enterprise-grade servers, and cutting-edge AI models. The company aims to dissolve the boundaries between simulation and reality across robotics, autonomous systems, and smart infrastructures.

During the computer graphics conference SIGGRAPH 2025, Nvidia made several significant announcements, centered around the concept of “physical artificial intelligence.” According to the company, this term refers to the integration of AI and computer graphics technologies to develop systems capable of interacting in the real world—whether it be robots, autonomous vehicles, or intelligent urban networks.

«AI enhances our modeling capabilities, and in turn, modeling makes AI smarter,» said Sanya Fidler, Nvidia’s VP of AI research. The company introduced a comprehensive ecosystem, ranging from new Blackwell architecture hardware to simulation platforms and specialized AI models designed for logical reasoning.

Nvidia has launched equipment intended to serve as a foundation for resource-intensive AI tasks. For data centers, they have prepared the Nvidia RTX PRO 6000 Blackwell Server Edition, a graphics accelerator in the widely-used 2U format, designed for standard corporate servers. Partners like Cisco, Dell, HPE, Lenovo, and Supermicro are already developing their solutions. As emphasized by Nvidia, these systems are intended to facilitate a transition away from traditional architectures solely reliant on CPUs to platforms focused on accelerated computing. These servers are claimed to offer up to a 45-fold increase in performance and an 18-fold boost in energy efficiency compared to CPU-only systems. The GPUs are equipped with fifth-generation tensor cores supporting the FP4 format, which, according to the company, enhances inference speed six times over the previous L40S model.

For desktop workstations, Nvidia announced two new compact cards: the Nvidia RTX PRO 4000 SFF Edition and the RTX PRO 2000 Blackwell. These are aimed at bringing AI acceleration to more compact and energy-efficient form factors for engineers, designers, and 3D visualization specialists. Compared to the previous generation, the RTX PRO 4000 SFF promises up to 2.5 times the AI performance while consuming the same 70 watts of power. Additionally, the RTX PRO 2000 is expected to perform 40% faster in computer-aided design (CAD) tasks. Nvidia anticipates that these innovations will be available for purchase by the end of this year.

This new hardware is designed to provide the computational power necessary to realize Nvidia’s vision of physical AI, which fundamentally involves creating highly realistic, physically accurate digital twins where AI systems, including robots, can safely learn through trial and error before operating in the real world.

«Computer graphics and AI are merging to fundamentally transform robotics,» underscores Rev Lebaredian, Nvidia’s VP for the Omniverse and simulation technologies.

The technological foundation is established through the Nvidia Omniverse and Isaac platforms. The company announced new libraries for Omniverse, including Omniverse NuRec, which enables the recreation of real-world environments based on sensor data using 3D Gaussian splatting techniques. Robotics simulation applications, Isaac Sim 5.0 and Isaac Lab 2.2, are now available as open source on GitHub and have already integrated new rendering methods.

A practical application example comes from Amazon Devices & Services, employing a simulation-first approach in what is termed zero-touch manufacturing. In this process, CAD models of new products are uploaded to Isaac Sim to generate over 50,000 synthetic images for AI training. These models subsequently control robotic arms that autonomously conduct quality inspections or integrate new products into the production line, relying solely on skills acquired through simulation, without any changes to the actual physical equipment. Technologies like the FoundationPose pose estimation model enable robots to recognize entirely new objects without prior training.

To ensure that AI is not only capable of seeing but also of thinking, Nvidia has expanded its model lineup. The corporate segment now includes the Nemotron family, with the introduction of the Nemotron Nano 2 and Llama Nemotron Super 1.5. These models are designed for AI agents to tackle complex multi-step tasks, ranging from customer service to cybersecurity. According to Nvidia, these models are highly efficient due to their hybrid architecture and NVFP4 format quantization. Companies such as CrowdStrike, Uber, and Zoom are already testing or preparing to implement them into their operations.

Specifically for physical AI, Nvidia created Cosmos Reason, a customizable vision-language model (VLM) with 7 billion parameters. This model aims to teach robots and computer vision systems how to interpret events in the physical world and act based on existing knowledge, understanding of physics laws, and commonsense reasoning. Applications include robotic task planning, automatic labeling of training data, and video analytics. For instance, Uber uses Cosmos Reason for analyzing the behavior of autonomous vehicles, while VAST Data and Milestone Systems utilize it for smart monitoring of transportation flows.

To bring these technologies into practical solutions for intelligent infrastructure, Nvidia consolidates multiple components into the Metropolis platform. It has received several updates, including the integration of the multimodal Cosmos Reason, new foundational computer vision models in the TAO Toolkit, and enhancements to Isaac Sim for generating rare training scenarios.

Partners are already finding practical applications for the platform. For example, Accenture and Belden are developing «smart virtual fences» modeled in Omniverse to enhance the safety of workers near industrial robots. DeepHow leverages the Metropolis VSS template to create a «smart all-knowing assistant»—a helper that transforms work instructions into visual guides. The brewing company Anheuser-Busch InBev claims that with this technology, it has cut the training time for new employees by 80%.

**Delegate routine tasks with BotHub!** The service is accessible without VPN, and Russian cards are accepted. [Click here](https://bothub.chat/?invitedBy=m_aGCkuyTgqllHCK0dUc7) to get 100,000 free caps and start working with neural networks right now!