Nvidia Unveils Agentic AI & Physical Robotics Models, Boosts Training Accuracy
Nvidia has unveiled significant advancements in the realm of artificial intelligence, introducing new agentic AI capabilities and groundbreaking physical robotics models. These developments, announced at SIGGRAPH 2025, mark a pivotal moment in the company’s commitment to enabling AI systems that can reason, plan, and interact with the physical world.
At the heart of Nvidia’s latest push is “agentic AI,” a paradigm shift towards AI systems that possess sophisticated reasoning and iterative planning abilities, allowing them to autonomously tackle complex, multi-step problems. This goes beyond traditional AI chatbots, enabling systems to analyze challenges, devise strategies, and execute tasks independently, promising enhanced productivity and operational efficiency across various industries. Nvidia is expanding its Nemotron and Cosmos model families to power these smarter AI agents, with new models like Nemotron Nano 2 and Llama Nemotron Super 1.5 offering improved reasoning accuracy and efficiency for enterprise applications. These models are designed to serve as the “brain” of AI agents, providing the core intelligence needed for complex workflows and real-world interactions.
A major highlight of Nvidia’s announcement is the substantial progress in “physical AI,” which focuses on empowering AI systems to perceive, reason, plan, and act within real-world environments. This initiative is particularly impactful for the rapidly evolving fields of robotics and autonomous systems. Central to this advancement is Cosmos Reason, a newly introduced 7-billion-parameter reasoning vision-language model (VLM). Designed specifically for robots and vision AI agents, Cosmos Reason enables these machines to understand complex instructions and plan actions by integrating memory, physics understanding, and common sense gleaned from training data. This allows robots to “reason” about what they see and determine the necessary steps for an embodied agent to take, making it invaluable for tasks such as data curation, robot planning, and video analytics.
Nvidia emphasizes that Cosmos Reason is trained using a combination of supervised fine-tuning and reinforcement learning, a methodology that has demonstrated notable performance improvements across key robotics and autonomous driving benchmarks. The company reports that post-training, the model’s performance on physical AI tasks improves by over 10%, with reinforcement learning contributing an additional 5%, achieving an average score of 65.7 across these benchmarks.
These advancements are not isolated; they are part of a comprehensive ecosystem designed to accelerate the development and deployment of physical AI solutions. Nvidia is rolling out new Omniverse libraries, including those for 3D Gaussian splatting for large-scale world reconstruction, and updating its Isaac Sim and Isaac Lab platforms for robust robot simulation. These tools allow developers to create physically accurate digital twins and generate synthetic data, crucial for safely training AI systems through trial and error before real-world deployment. Furthermore, Nvidia is bolstering its infrastructure with new Blackwell-powered RTX Pro Servers, specifically engineered to handle the intensive computational demands of these advanced AI workloads. These servers, available in various configurations through partnerships with leading providers like HPE, aim to make high-performance AI inference accessible for enterprise and industrial applications.
The announcements at SIGGRAPH 2025 underscore the growing convergence of artificial intelligence and computer graphics, with Nvidia positioning itself at the forefront of this transformation. By enabling AI agents to reason more intelligently and providing robust tools for physical AI training, Nvidia is laying the groundwork for a future where intelligent machines can seamlessly understand and operate within our complex physical world.