It wasn’t as wild as an episode of ‘Wacky Races’, but Nvidia drily showcased its interest in winning the race for autonomous driving glory.
The tech titan used the global NeurIPS AI conference to introduce a set of open-source models, datasets, and tools that the company says could accelerate research across fields ranging from autonomous driving to speech recognition and AI safety.
The releases include what Nvidia describes as the world’s first open, industry-scale reasoning vision-language-action (VLA) model for mobility, alongside expanded digital AI offerings under its Nemotron and NeMo ecosystems.
The push underscores a broader trend in the AI sector: researchers and developers increasingly expect transparency and open tooling in order to evaluate system performance, reproduce resultsm and build their own customized models. The firm’s latest announcements position it more firmly within that open ecosystem.
A new ranking from Artificial Analysis, an independent benchmarking organization, reflected this shift by placing the company’s Nemotron line of models and datasets among the most open in the industry. The score was based on licensing terms, data transparency, and the availability of technical documentation.
Autonomous driving
The headline announcement was Nvidia DRIVE Alpamayo-R1 (AR1), described as the world’s first open reasoning VLA model built for autonomous vehicle (AV) research. The model merges chain-of-thought-style reasoning with path planning, a combination intended to help vehicles interpret complicated real-world scenarios and make higher-confidence decisions.
Traditional AV systems have long struggled with highly variable urban environments, where construction zones, unpredictable pedestrians, or blocked bike lanes require context-aware judgment. Nvidia says AR1 helps address this by decomposing scenes into steps, assessing multiple potential trajectories, and selecting a course of action informed by both sensor data and reasoning traces.
The company offered an example in which an AV navigating a busy street could use AR1’s reasoning to modify its path based on pedestrian density near a bike lane. The firm said the model’s structure enables vehicles to “incorporate reasoning traces — explanations on why it took certain actions” and apply them to future decisions.
The model is intended for non-commercial research use and will be available on GitHub and Hugging Face, with a subset of the training data included in Nvidia’s open Physical AI datasets. To evaluate AR1, the company also released AlpaSim, an open-source simulation framework.
As the industry inches toward Level 4 autonomy, such tools could help academic labs and startups experiment more safely and at lower cost.
Physical AI
Beyond driving, Nvidia expanded its Cosmos family of world and policy models, releasing new materials aimed at robotics, simulation, and embodied AI. The Cosmos Cookbook, a guide that covers data curation, synthetic data generation, and post-training techniques, is now available to developers.
Several new Cosmos-based models were showcased, including LidarGen, which can generate synthetic lidar data for AV simulation; Omniverse NuRec Fixer, which cleans artifacts in neurally reconstructed scenes; Cosmos Policy, which converts video foundation models into robotic control policies; and ProtoMotions3, a GPU-accelerated framework for training physically simulated digital humans.
These tools reduce the time and cost required to produce realistic training environments for robots and AV systems. Reliable synthetic data has become a critical ingredient for advanced physical AI, especially as real-world data collection grows more expensive, restrictive, and safety-sensitive.
A number of ecosystem partners — including 1X, Figure AI, Gatik, and ETH Zurich — are already integrating Cosmos world foundation models into their pipelines, signaling wider adoption beyond Nvidia’s own research groups.
Speech, safety, and RL development
On the digital AI side, Nvidia released new speech recognition models and expanded its suite of tools for AI safety and reinforcement learning. MultiTalker Parakeet and Sortformer address multi-speaker recognition and diarization, enabling models to understand fast-paced or overlapping conversations.
The company also introduced Nemotron Content Safety Reasoning, which applies reasoning techniques to enforce custom safety policies. An accompanying synthetic dataset for unsafe audio scenarios aims to help organizations build guardrails that work across both text and audio.
Reinforcement learning developers gained two new open-source resources: NeMo Gym, which provides turnkey RL environments tailored to language model training, and the NeMo Data Designer Library, now open under Apache 2.0, which gives developers an end-to-end pipeline for generating and validating synthetic datasets.
These additions respond to a growing need for AI systems that can operate safely, handle real-time speech inputs, and learn from structured reward environments — capabilities increasingly sought by enterprises in sectors ranging from cybersecurity to automation.
Last month, the Trump administration was weighing whether to let Nvidia sell H200 AI chips to China, pitting national security worries against a massive chip market.