Nvidia has unveiled more than six AI models specifically designed for autonomous systems, targeting applications such as self-driving cars.
These algorithms were released under open-source licenses and made their debut at the CES technology exhibition in Las Vegas. The launch also includes a suite of development tools and a new robotics computing module named Jetson T4000.
Autonomous Vehicle Software
The standout among Nvidia’s newly introduced open-source AI models is Alpamayo 1 (shown above), a 10-billion-parameter visual-language-action (VLA) model capable of generating driving trajectories from camera feeds captured by autonomous vehicles.
Alpamayo 1 features a chain-of-thought reasoning mechanism, enabling it to break down navigation tasks into smaller, interpretable steps. According to Nvidia, this approach offers two key advantages: first, it allows the model to explain each stage of its decision-making process, making it easier to evaluate the logic behind navigation choices. Second, the step-by-step reasoning enhances performance in complex driving scenarios.
While not intended to run directly on autonomous vehicles, Alpamayo 1 is designed to help developers train navigation models for such systems. Nvidia states the algorithm is ideal for evaluating the reliability of self-driving software. The company plans to release larger versions of the Alpamayo series in the future to support broader reasoning applications.
"Alpamayo brings reasoning capabilities to autonomous vehicles, enabling them to drive safely in complex environments and explain their decisions—foundational elements for safe, scalable autonomy," said Jensen Huang, CEO of Nvidia.
Alpamayo 1 is accompanied by three new additions to Nvidia’s existing Cosmos family of world foundation models. Like Alpamayo 1, these models are tailored for developing software for autonomous vehicles and can also support other automated systems, including industrial robots.
The first two, Cosmos Transfer 2.5 and Cosmos Predict 2.5, are designed to generate synthetic training data for robotic AI software in the form of simulated video sequences. For instance, Cosmos Transfer 2.5 can create footage of industrial robots operating in automotive factories. Cosmos Predict 2.5 offers similar functionality with the added ability to simulate future object behaviors—users can upload an image of a bus and request predictions of its position five seconds later.
The third addition, Cosmos Reason 2.0, enables robots to analyze visual inputs from their environment and perform autonomous actions. It powers Isaac GR00T N1.6, another new model launched today. Isaac GR00T N1.6 is also a VLA model but optimized for humanoid robots rather than self-driving vehicles. Nvidia trained the algorithm on a dataset comprising sensor data from bimanual, semi-humanoid, and full humanoid robots.
"Salesforce, Milestone, Hitachi, Uber, VAST Data, and Encord are using Cosmos Reason to power AI agents for transportation and workplace productivity," wrote Kari Briski, Vice President of Generative AI Software at Nvidia, in a blog post. "Franka Robotics, Humanoid, and NEURA Robotics are leveraging Isaac GR00T to simulate, train, and validate new robotic behaviors before scaling to production."
Nvidia’s robotics-focused models were launched alongside two more general-purpose model families: Nemotron Speech and Nemotron RAG. The former features a speech recognition model claimed to outperform comparable alternatives by up to 10x. Nemotron RAG includes embedding and re-ranking models.
Embedding models convert data into mathematical representations that AI systems can understand. Re-ranking is a critical step in retrieval-augmented generation (RAG) workflows. After an AI system retrieves documents relevant to a query, the re-ranking model highlights the most pertinent ones for response generation.
Open-Source Development Tools
Nvidia's AI models come paired with three open-source development tools. The first, AlpaSim, allows developers to build customized simulation environments for training autonomous driving models. It supports fine-tuning variables such as traffic conditions and sensor configurations. To increase difficulty, developers can introduce sensor noise to test how well their AI models handle corrupted or inaccurate data.
Nvidia also introduced Isaac Lab-Arena, a second simulation framework aimed at simplifying the training of robotic AI models. The platform supports popular third-party benchmarks like Robocasa, commonly used for assessing home robot performance, enabling developers to benchmark their models effectively.
The third tool, OSMO, helps software teams manage simulation workloads. Acting as a workflow orchestrator, OSMO supports various AI development pipelines, including synthetic data generation and model training. Nvidia notes that OSMO can coordinate tasks across public cloud platforms and local developer workstations.
New Hardware
A new computing module from Nvidia, the Jetson T4000, is now available for manufacturers to power their robotic systems. Built on the company’s Blackwell GPU architecture, the module can run AI-driven navigation software for factory-floor robotics, for example.
The Jetson T4000 comes with 64GB of memory and delivers up to 1,200 teraflops (TFLOPS) of compute performance when processing FP4 data—making it four times faster than Nvidia’s previous-generation robotics module. The Jetson T4000 will be offered at $1,999 per unit for customers purchasing at least 1,000 units.
