The Nvidia World Headquarters in Santa Clara.iStock Photos
NVIDIA is expanding the frontiers of AI research with a bold move.
The company on Monday announced the launch of open physical and digital AI models that could reshape autonomous vehicles, robotics, and speech processing.
The company unveiled these breakthroughs at NeurIPS, one of the world’s top AI conferences, signaling a new era for open-source AI development.
Among the highlights is Alpamayo-R1 (AR1), the world’s first open reasoning vision-language-action (VLA) model for autonomous driving.
Designed to combine chain-of-thought reasoning with path planning, AR1 helps vehicles navigate complex scenarios with human-like judgment.
“AR1 accomplishes this by breaking down a scenario and reasoning through each step,” NVIDIA said.
The model evaluates possible trajectories and uses contextual data to select the safest route, making it capable of handling intersections crowded with pedestrians, double-parked vehicles, or upcoming lane closures.
Open access is central to NVIDIA’s strategy. AR1, built on NVIDIA Cosmos Reason, can be customized by researchers for non-commercial applications.
Reinforcement learning post-training has shown significant improvements in the model’s reasoning capabilities compared to its pretrained version.
Reasoning drives autonomy
AR1 is available on GitHub and Hugging Face, alongside a subset of training data in the NVIDIA Physical AI Open Datasets. Researchers can also leverage the AlpaSim framework to evaluate model performance.
Beyond AR1, NVIDIA’s Cosmos platform offers a suite of tools for physical AI development.
These include LidarGen, which generates lidar data for AV simulation; Omniverse NuRec Fixer for cleaning up neural reconstructions; Cosmos Policy to create robot behavior rules; and ProtoMotions3, which trains humanoid robots in realistic, simulated environments.
Developers and researchers worldwide are already experimenting with these models. NVIDIA ecosystem partners like Voxel51, 1X, Figure AI, Foretellix, Gatik, Oxa, PlusAI, and X-Humanoid are applying Cosmos foundation models in autonomous driving and robotics projects.
ETH Zurich researchers are also using Cosmos for 3D scene creation, presenting their findings at NeurIPS.
Digital AI expansion
On the digital front, NVIDIA is enhancing its Nemotron toolkit. Highlights include MultiTalker Parakeet, a multi-speaker automatic speech recognition model, and Sortformer, which performs real-time speaker diarization.
Nemotron also introduces reasoning-based AI safety models and synthetic datasets for reinforcement learning and domain-specific AI development.
Other notable updates include Audio Flamingo 3, a large audio language model capable of reasoning across speech, music, and sound, andMinitron-SSM, which compresses hybrid models without sacrificing performance.
NVIDIA researchers also showcased Jet-Nemotronand Nemotron-Flash, optimized for efficient language model inference and latency.
“Prolonged reinforcement learning, or ProRL, is a technique that extends model training over longer periods. This methodology results in models that consistently outperform base models for reasoning,” NVIDIA noted.
With these announcements, NVIDIA continues to push open-source AI to new heights, offering tools that empower researchers, developers, and autonomous system innovators alike.
The company’s efforts were recognized by Artificial Analysis, which rated NVIDIA’s Nemotron family among the most open in the AI ecosystem.
The Blueprint