NVIDIA has unveiled a major expansion of its open-source AI initiatives at NeurIPS 2024, reinforcing its role as a leader in advancing both digital and physical AI. The company introduced new models, tools, and datasets designed to accelerate research across autonomous vehicles, robotics, and AI safety, with a strong emphasis on open collaboration. A centerpiece of the announcement is Alpamayo-R1 (AR1), the world’s first open reasoning vision language action (VLA) model tailored for autonomous driving. Built on NVIDIA’s Cosmos Reason foundation, AR1 integrates chain-of-thought reasoning with path planning, enabling self-driving systems to make human-like decisions in complex, real-world scenarios—such as navigating pedestrian-heavy intersections or responding to double-parked vehicles. By breaking down driving situations step by step and generating explainable reasoning traces, AR1 improves decision-making transparency and safety, a critical step toward achieving level 4 autonomy. AR1 is available on GitHub and Hugging Face, and a subset of the training and evaluation data is included in the NVIDIA Physical AI Open Datasets. The company also released AlpaSim, an open-source framework to evaluate AR1’s performance, further supporting reproducibility and innovation in AV research. This open approach aligns with NVIDIA’s broader commitment to open source, which was recently recognized by the Artificial Analysis Openness Index, ranking the NVIDIA Nemotron family as one of the most open in the AI ecosystem due to permissive licenses, data transparency, and detailed technical documentation. To help developers build and customize physical AI systems, NVIDIA launched the Cosmos Cookbook—a comprehensive guide with step-by-step recipes, inference examples, and post-training workflows. The cookbook supports data curation, synthetic data generation, and model evaluation, enabling researchers and engineers to adapt Cosmos-based models for diverse applications. Notable examples include LidarGen, which generates realistic LiDAR data, and policy models trained in NVIDIA Isaac Lab and Isaac Sim, which can be used to fine-tune GR00T N models for robotics. Partners like Voxel51, Figure AI, Oxa, and researchers from ETH Zurich are already leveraging these tools for projects ranging from 3D scene generation to humanoid robot control. In the digital AI space, NVIDIA expanded its Nemotron toolkit with new multi-speaker speech models, a reasoning-capable model, and datasets focused on AI safety. The company also introduced tools for generating high-quality synthetic data, essential for reinforcement learning and domain-specific model training. Ecosystem partners including CrowdStrike, Palantir, and ServiceNow are using these tools to build secure, specialized agentic AI systems. NVIDIA researchers are presenting over 70 papers, talks, and workshops at NeurIPS, covering breakthroughs in language models, medical AI, and autonomous systems. The company’s push into physical AI reflects a strategic shift, with CEO Jensen Huang and Chief Scientist Bill Dally emphasizing that the next frontier of AI lies in machines that perceive and act in the real world. As physical AI becomes central to robotics, transportation, and industry, NVIDIA aims to be the foundational “brain” for these systems, powered by its advanced GPUs and open, accessible tools. These developments mark a significant step toward democratizing access to cutting-edge AI technology and accelerating innovation across research and industry.