Nvidia debuts new robotics and self-driving technology at CES 2026

Key Concepts

• Omniverse: A platform for 3D design collaboration and physically accurate simulation.
• Cosmos: A foundation model used to generate physically based, synthetic data for AI training.
• AlphaMayo: Nvidia’s end-to-end autonomous vehicle AI, trained with both real and synthetic data.
• Ruben: Nvidia’s new supercomputer, comprised of 1152 GPUs, designed for AI development.
• Synthetic Data: Artificially generated data used to supplement or replace real-world data for AI training.
• End-to-End AI: An AI system trained directly from raw input (e.g., camera images) to final output (e.g., vehicle control).
• Physical AI: AI systems that interact with and understand the physical world.

Advancements in Physical AI and Autonomous Vehicles – A Detailed Overview

Introduction of Robotics and Simulation Capabilities

The presentation began with a demonstration of robots trained within Nvidia’s Omniverse platform. Several robotic platforms were showcased, including Neurobot, Aubot (developed by LG), Surf Robot (connected to Uber Eats), and robots from Agility Robotics and Boston Dynamics. The ability to create realistic simulations within Omniverse was highlighted, allowing robots to “learn” in a virtual environment before deployment in the real world. A demonstration showed the conversion of a 3D scene into an architectural rendering and then a navigable video, showcasing the platform’s versatility. Brev was mentioned as a tool enabling shared access to simulation environments. A playful interaction with a virtual assistant (“Potato”) demonstrated the ability to issue commands within the simulated environment.

The Need for Synthetic Data and the Introduction of Cosmos

A core argument presented was the limitation of relying solely on real-world data for training AI, particularly for complex tasks like autonomous driving. Collecting sufficient real-world data is slow, expensive, and often inadequate. The solution, according to Nvidia, is synthetic data. Cosmos, a new foundation model, was introduced as a key component in generating this data. Cosmos takes the output of traffic simulators and generates physically plausible, surround video that AI can learn from. This addresses the challenge of the physical world’s diversity and unpredictability. As stated, “The chat GPT moment for physical AI is nearly here.”

AlphaMayo: Nvidia’s Autonomous Vehicle AI

The centerpiece of the presentation was the unveiling of AlphaMayo, described as “the world’s first thinking reasoning autonomous vehicle AI.” AlphaMayo is an end-to-end AI, meaning it’s trained directly from camera input to vehicle actuation. Training data consists of both real-world miles driven by humans and vast amounts of synthetic data generated by Cosmos, supplemented by hundreds of thousands of carefully labeled examples. A five-year partnership with Mercedes-Benz was crucial in developing AlphaMayo. The vision is for widespread autonomous vehicle adoption, with options for robo-taxis, personal ownership, or manual driving, but with all cars possessing autonomous capabilities. The model powering this is AlphaMayo, with the application layer being the Mercedes-Benz system.

Deployment Timeline and Continuous Updates

AlphaMayo is slated for initial deployment in Nvidia-powered Mercedes-Benz vehicles in Q1 2024, followed by Europe in Q2, the United States in Q1, and Asia in Q3/Q4. Crucially, Nvidia plans to continuously update AlphaMayo with newer versions, ensuring ongoing improvement and adaptation.

Introducing Ruben: The AI Supercomputer

To support the development of these advanced AI models, Nvidia unveiled Ruben, a massive supercomputer consisting of 1152 GPUs across 16 racks. Each rack contains 72 Vera Rubin GPUs, each comprised of two GPU dies. The assembly process takes two hours and features a highly efficient design with minimal cabling.

Industry Perspective and Future Outlook

An industry analyst commented on Nvidia’s aggressive move into the autonomous vehicle space, contrasting it with a more cautious approach taken in the past. They highlighted the significance of Nvidia’s end-to-end platform and positioned Nvidia and Tesla as the leading players in physical AI. The analyst stated, “I think this was Jensen much more on the offensive…planting a flag when it comes to physical AI.”

The presentation concluded with a strong assertion that advancements in robotics are happening faster than anticipated, with humanoid robots potentially becoming commonplace in US households within 12-18 months. The speaker emphasized, “This is not 5 10 years away. This is something that's actually coming a lot sooner.” The ability to build applications using pre-trained language models is accelerating this progress.

Technical Terms & Concepts:

• GPU (Graphics Processing Unit): Specialized electronic circuits designed to rapidly manipulate and display computer graphics. Crucial for AI training and inference.
• Foundation Model: A large AI model pre-trained on a massive dataset, capable of being adapted to a wide range of downstream tasks.
• Actuation: The process of controlling a device or system, in this case, the control mechanisms of a vehicle.
• Physically Based Rendering (PBR): A rendering technique that simulates the physical properties of materials and light to create realistic images.
• End-to-End Learning: Training a model directly from raw input to output without intermediate steps or hand-engineered features.

Logical Connections:

The presentation flowed logically from demonstrating the need for advanced simulation (Omniverse) to the solution of generating synthetic data (Cosmos) to the application of this data in a fully autonomous system (AlphaMayo). The introduction of Ruben underscored Nvidia’s commitment to providing the computational infrastructure necessary for these advancements. The analyst’s commentary provided external validation of Nvidia’s strategic direction.

Data & Statistics:

• 1152 GPUs in the Ruben supercomputer.
• 72 Vera Rubin GPUs per rack.
• Two GPU dies per Vera Rubin GPU.
• Hundreds of thousands of labeled examples used for AlphaMayo training.
• Deployment timeline: Q1 2024 (Europe), Q2 2024 (US), Q3/Q4 2024 (Asia).

Conclusion:

Nvidia presented a compelling vision for the future of AI, particularly in the realm of autonomous vehicles and robotics. The combination of advanced simulation capabilities (Omniverse), synthetic data generation (Cosmos), and a powerful end-to-end AI (AlphaMayo), supported by substantial computational resources (Ruben), positions Nvidia as a major force in the rapidly evolving field of physical AI. The emphasis on continuous updates and a clear deployment timeline suggests a commitment to bringing these technologies to market quickly and effectively.