NVIDIA provides an end-to-end stack for autonomous vehicle development—from sensor data curation and synthetic data generation to model training, closed-loop simulation, and production-grade in-vehicle compute.
Transform raw sensor data into high-quality training datasets with the NVIDIA Physical AI Data Factory Blueprint. NVIDIA Cosmos™ Curator and Cosmos Dataset Search automate data curation and rare scenario retrieval at scale. The data fine-tunes NVIDIA Alpamayo open VLA reasoning models and AlpaGym runs closed-loop RL post-training at GPU scale.
Accelerate AV development with high-fidelity, scalable simulation workflows. Reconstruct real-world driving logs into interactive simulation with NVIDIA Omniverse NuRec, generate photorealistic synthetic scenarios with Cosmos-Dreams, and run closed-loop policy evaluation at GPU scale with AlpaSim.
DRIVE AGX™ is an automotive-grade compute platform delivering industry-leading performance. High-performance DRIVE AGX compute, paired with trusted NVIDIA DRIVE AGX Orin™ and DRIVE AGX Thor™ ecosystem partners, accelerates production deployment. DRIVE AGX is powered by the DriveOS™ SDK featuring NVIDIA DriveWorks, CUDA®, TensorRT™, NvMedia, and NvStreams.
The NVIDIA DRIVE AGX developer kit is the reference compute platform for the DRIVE Hyperion™ architecture, providing production-equivalent in-vehicle hardware for integrating and testing AV software stacks before vehicle-level deployment.
NVIDIA DriveOS is the safety-certified automotive operating system for DRIVE AGX, providing sensor abstraction, compute scheduling, and middleware integration for automotive-grade AV development.
NVIDIA Cosmos Curator filters, annotates, and deduplicates large amounts of sensor data necessary for autonomous vehicle development. It taps into NVIDIA Cosmos Reason VLM for multimodal reasoning and shortens data processing pipelines from months to days.
NVIDIA Cosmos Dataset Search instantly searches and retrieves targeted scenarios from massive training datasets, powered by Cosmos-Embed NIM for semantic search across billions of clips in seconds.
NVIDIA CoC Auto-Labeling Pipeline automatically generates Chain of Causation reasoning labels for driving clips, removing manual annotation and accelerating dataset preparation for model training.
NVIDIA Omniverse NuRec uses Gaussian-based methods to reconstruct and render interactive simulation from real-world driving data. Its 3DGUT core combines the speed of Gaussian splatting with the photorealism of ray-tracing for physically accurate AV simulation. NVIDIA InstantNuRec accelerates reconstruction by generating a 3D Gaussian splat in a single forward pass, reducing NuRec training iterations by up to 25% on a single GPU.
NVIDIA Cosmos Transfer is a world foundation model that generates photorealistic synthetic driving data from structured inputs, including HD Maps, lidar depth, and text prompts. Transfer produces multi-view consistent video across weather, lighting, and environmental variations for AV perception and planning model training.
NVIDIA Omniverse NuRec generative models improve reconstruction quality and scene diversity. Fixer removes flickering artifacts from rendered novel views. NVIDIA Harmonizer normalizes lighting across weather and time-of-day variations. NVIDIA Asset Harvester extracts 3D Gaussian objects from partial views.
NVIDIA Alpamayo 2 Super is an open 34B VLA reasoning foundation model with RL post-training, flexible multi-camera support, and navigation guidance. Post-training scripts for SFT and RL fine-tuning on proprietary fleet data are available on GitHub under Apache 2.0.
NVIDIA AlpaGym is the first modular RL framework for training AV policy models at GPU scale, running models through continuous decision and observation cycles to expose compounding errors that static datasets miss.
NVIDIA AlpaSim is an open-source closed-loop AV simulation framework. Its microservice architecture assigns rendering, physics, traffic behavior, and policy execution to separate GPU resources, with support for NVIDIA Omniverse NuRec and NVIDIA Cosmos-Dreams as rendering backends.
NVIDIA Omniverse NuRec reconstructs captured driving scenes into interactive 3D Gaussian splat environments for regression testing and safety evaluation, with sub-25 ms photoreal frame playback.
NVIDIA Cosmos-Dreams renders photorealistic camera frames conditioned on live policy actions for continuous novel scenario generation in closed-loop simulation, at up to 54 fps on 1x GB300 and 30 fps on RTX 6000.