1. [Home ](/) [DRIVE](/drive) Downloads # NVIDIA Automotive Downloads **For AV Infrastructure / Simulation Downloads:** This page provides access to datasets, models, code, containers, and documentation for AV developers’ [data factory](https://www.nvidia.com/en-us/solutions/autonomous-vehicles/ai-training/) (data ingestion and processing, data search and curation, synthetic data generation), [training](https://www.nvidia.com/en-us/solutions/autonomous-vehicles/ai-training/), and [simulation](https://developer.nvidia.com/drive/simulation) workflows. Use the links below to access public GitHub repositories, Hugging Face models or datasets, NGC containers, documentation, and support channels. **For Embedded Downloads:** This page provides access to the NVIDIA DRIVE® SDK for developers using DRIVE hardware. See [Automotive Hardware](/drive/agx) and [Automotive Software](/drive/driveos) for more details. For support, please post any queries or issues in the [Forums](https://forums.developer.nvidia.com/c/autonomous-vehicles/517). Please note that the download requires membership to the [NVIDIA DRIVE AGX™ SDK Developer Program](/drive-agx-program) for DRIVE AGX, and to the [NVIDIA DRIVE PX 2 SDK Developer Program](/drive-px2-program) for DRIVE PX 2. By downloading or using the software and materials, you agree to the [NVIDIA Automotive Software License Agreement for Development on NVIDIA SoC](https://developer.nvidia.com/downloads/licenses/NVIDIA-Automotive-Software-License-for-Development-with-NVIDIA-SoC). The rows below are the raw result records for this component — one row per record. Column names are the source index fields; the full data is also downloadable in the links that follow. | date | links | title | markdown | products | subtitle | | --- | --- | --- | --- | --- | --- | | 1793232000 | {"url" => "https://docs.nvidia.com/nurec/", "text" => "More Information"} | Omniverse NuRec | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/NVIDIA/NuRec' target='_blank' class='icon icon-page'>GitHub Repository</a></li> </ul> </div> </div> <div data-markdown="1"> Use NVIDIA Omniverse™ NuRec, a set of APIs and libraries for Gaussian-based reconstruction and rendering of interactive simulations from real-world data. Omniverse NuRec combines reconstruction libraries and containers, rendering APIs, and generative AI models for physical AI simulation and synthetic data generation, including autonomous driving and robotics. Start with the NVIDIA/NuRec GitHub repository for SDK source, container workflows, related models and datasets, and documentation. ##### Supported hardware: * Minimum GPU: Single NVIDIA GPU with CUDA support (version 12.8 or higher); more than 24 GB memory, with more than 48 GB recommended. * Host OS: Linux x86_64 * Other dependencies: Docker with NVIDIA Container Toolkit; NCore-format data for reconstruction. Source builds use the NuRec development container with Python 3.13, Bazel, and an r580-or-newer driver. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVIDIA/NuRec/issues' target='_blank'>Omniverse NuRec GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/omniverse/platform/nurec/' target='_blank'>NVIDIA Developer Forum — Omniverse NuRec</a></li> </ul> ##### Omniverse NuRec 26.10 includes the following benefits: * Modular, GPU-accelerated SDK libraries let developers customize reconstruction and rendering components or integrate them into an existing pipeline. * Camera and ray-based LiDAR rendering let developers reuse a reconstructed scene across sensor configurations and simulated vehicle poses. * Reusable OpenUSD scenes support synthetic data generation and repeated evaluation in AV and robotics simulators. * An AV benchmark workflow connects reconstruction, rendering, evaluation, and scene export with task timings and image, object, and LiDAR metrics. <a href='https://github.com/NVIDIA/NuRec' target='_blank'>Read more</a> about Omniverse NuRec </div> | Omniverse NuRec | Version 26.10 \| Release date 2026/10/29 | | 1793145600 | {"url" => "https://github.com/NVIDIA/NuRec", "text" => "More Information"} | NCore | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/NVIDIA/ncore' target='_blank' class='icon icon-page'>Github Repository</a></li> <li><a href='https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles-NCore' target='_blank' class='icon icon-download'>NCore-converted Physical AI AV dataset</a></li> </ul> </div> </div> <div data-markdown="1"> Use NVIDIA NCore — data representations, APIs, and tools to support data-driven neural reconstructions with a focus on robotics and autonomous-vehicle data. NCore provides a standardized data format for sensor recordings used by NuRec reconstruction and simulation workflows. ##### Supported hardware: * Minimum GPU: not required for ingestion or conversion; GPU is used downstream by NuRec, Asset Harvester, etc. * Host OS: any platform supported by the nvidia-ncore PyPI package (Linux + macOS development environments tested). * Other dependencies: **pip install nvidia-ncore**; Python; converters transform external datasets into a common representation (see <a href='https://nvidia.github.io/ncore' target='_blank'>nvidia.github.io/ncore</a> for the current supported list). ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVIDIA/ncore/issues' target='_blank'>NCore GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/omniverse/platform/nurec/' target='_blank'>NVIDIA Developer Forum — Omniverse NuRec</a></li> </ul> ##### NCore includes the following benefits: * Standardized multi-sensor data format lets you convert camera, LiDAR, calibration, and pose data from any platform into a single NuRec-ready representation — eliminate proprietary data pipeline engineering. * GPU-accelerated camera and LiDAR intrinsic models built in give downstream renderers and reconstructors the same physically-accurate sensor models, removing a class of subtle simulation bugs. * Use custom camera, LiDAR, and external distortion models, with parameters that can be saved and loaded in NCore. * Converters transform external datasets into the common NCore representation — drop-in adapters maintained as the supported-format list grows (see the NCore docs for the current set). * Pip-installable Python package with full reference documentation on nvidia.github.io/ncore makes adoption a one-liner — no compilation, no container. </div> | Omniverse NuRec | Version 19.8.5 \| Release date 2026/10/28 | | 1792540800 | {"url" => "https://www.nvidia.com/en-us/ai/cosmos/", "text" => "More Information"} | Cosmos Evaluator | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/nvidia-cosmos/cosmos-evaluator' target='_blank' class='icon icon-page'>GitHub Repository</a></li> </ul> </div> </div> <div data-markdown="1"> Use NVIDIA Cosmos Evaluator to automate evaluation and grading of synthetic video outputs generated by Cosmos models, including quality and condition checks for Cosmos Data Factory workflows. The four reference checks are Hallucination, Objects Consistency, Attribute Verification, and Physical Plausibility. ##### Supported hardware: * GPU requirements: NVIDIA GPU with CUDA for the default Objects and Physical Plausibility configurations; NVIDIA driver 570.124.06+ for GPU containers. Hallucination and Attribute Verification checker containers run on CPU. * Host OS: Linux; deployable as REST API microservices or as Python API calls. * Other dependencies: Python 3.10+, Docker 28.1.1+, Git LFS 3.0+, and NVIDIA Container Toolkit 1.16.2+ for GPU containers. Attribute Verification uses configured LLM/VLM endpoints; Physical Plausibility requires access to a compatible checkpoint. Apache License 2.0. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/nvidia-cosmos/cosmos-evaluator/issues' target='_blank'>Cosmos Evaluator GitHub Issues</a></li> <li><a href='https://discord.com/channels/827959428476174346/1469128459061301270' target='_blank'>NVIDIA Developer Discord — Cosmos channel</a></li> </ul> ##### Cosmos Evaluator includes the following benefits: * Automate dataset-scale evaluation with reusable AI judges that apply domain-specific rubrics across multiple criteria. * Screen generated videos for motion hallucinations, object mismatches, incorrect scene attributes, and implausible motion before downstream use. * REST API and Python API surfaces give you both microservice deployment for CI/CD-style quality gating and direct integration into research notebooks. * Train, fine-tune, or calibrate the Physical Plausibility checker with the included notebook to match your task and choose pass/fail thresholds. * The Arbitrator coordinates checks across camera streams, runs them in dependency order, and records results through the Database REST API and PostgreSQL. <a href='https://github.com/nvidia-cosmos/cosmos-evaluator' target='_blank'>Read more</a> about Cosmos Evaluator </div> | Cosmos | Version 26.10 \| Release date 2026/10/21 | | 1792540800 | {"url" => "https://www.nvidia.com/en-us/ai/cosmos/", "text" => "More Information"} | Cosmos Data Factory | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/NVIDIA/cosmos-data-factory' target='_blank' class='icon icon-page'>GitHub Repository</a></li> </ul> </div> </div> <div data-markdown="1"> Use NVIDIA Cosmos Data Factory, an orchestration framework for building synthetic image and video datasets for physical AI. Cosmos Data Factory connects data preparation, generation, annotation, and evaluation into configurable workflows for physical AI development. ##### Supported hardware: * GPU requirements: Workflow dependent. The AV reference setup allocates NVIDIA H100 80 GB GPUs; its WSM generation example uses four GPUs during generation, including two on the same node for Cosmos3-Nano. * Host OS: Linux; MicroK8s for local deployment or managed Kubernetes services for cloud deployment. * Other dependencies: Python 3.10+, Apache Airflow, Docker, Helm, and cloud storage services. AV workflows also use Milvus and a separately installed FFmpeg build. Apache License 2.0. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVIDIA/cosmos-data-factory/issues' target='_blank'>Cosmos Data Factory GitHub Issues</a></li> <li><a href='https://discord.com/channels/827959428476174346/1469128459061301270' target='_blank'>NVIDIA Developer Discord — Cosmos channel</a></li> </ul> ##### Cosmos Data Factory includes the following benefits: * Curate and retrieve relevant driving clips with Cosmos Curator and Cosmos Dataset Search, using text, video, or embedding queries. * Generate weather and lighting variants with Cosmos3-Nano, using RGB inputs or optional World Scenario Model (WSM) controls to guide scene structure. * Automatically annotate source and generated videos with object detections, tracks, captions, visual question-answer pairs, and reasoning labels, then export them for training. * Apply configurable Cosmos Evaluator checks to generated outputs and retain validation and evaluation results in output manifests. * Launch and monitor reusable workflows through the Airflow UI or REST API, with execution options that vary by workflow. Autonomous Vehicle Dataset Generation uses Airflow orchestration and Kubernetes execution. ##### Autonomous Vehicle Dataset Generation: Autonomous Vehicle Dataset Generation is the AV-specific reference workflow within Cosmos Data Factory. It generates targeted post-training datasets to help improve models on scenarios where they underperform, chaining functionalities relevant to the AV use case: 1. Embedding computation against a large video corpus using Cosmos Embed, to enable efficient search across the corpus. 2. Video retrieval of a targeted subset of the corpus using Cosmos Dataset Search, based on embedding similarity to a sample on which the model underperforms. 3. Video generation of synthetically-varied samples by Cosmos 3 to add to the targeted subset, based on selected seed videos from the subset together with user text prompts. 4. Evaluation of the physical plausibility of the generated videos using Cosmos Evaluator as a quality check. </div> | Cosmos | Version 1.0 \| Release date 2026/10/21 | | 1792540800 | {"url" => "https://github.com/NVlabs/alpamayo-recipes/", "text" => "More Information"} | CoC Autolabeling Pipeline | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/NVlabs/alpamayo-coc-autolabeler' target='_blank' class='icon icon-page'>GitHub Repository</a></li> </ul> </div> </div> <div data-markdown="1"> Use the NVIDIA Alpamayo CoC Auto-Labeling Pipeline — an open autolabeling pipeline that derives meta-actions and Chain-of-Causation reasoning labels connecting causal factors to ego-vehicle behavior. The pipeline reads trajectory data to produce per-clip high-level motion labels, identifies keyframes at ego action transitions, then runs a VLM over those keyframes to emit causally grounded reasoning traces in YAML. Version 2.0 introduces Reasoning Scene Graphs (RSG), a versatile new format that represents reasoning with structured graphs. ##### Supported hardware: * Minimum GPU: NVIDIA GPU required when using local inference with Qwen VLM. GPU optional when using hosted VLM APIs such as GPT-5 / GPT-5.5 * Recommended GPU: NVIDIA A100 or H100 for local Qwen inference; CUDA 12.8 with NVIDIA driver ≥535 (Ampere) or ≥545 (Hopper). * Host OS: Linux; minimal CPU-only setup supported (8 cores, 8 GB RAM) when using hosted APIs. * Other dependencies: vLLM 0.17.1 for local Qwen serving; trajdata framework for Physical AI AV dataset integration ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVlabs/alpamayo/issues' target='_blank'>Alpamayo GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/autonomous-vehicles/alpamayo/766' target='_blank'>NVIDIA Developer Forum — Alpamayo</a></li> </ul> ##### CoC Labeling Pipeline includes the following benefits: * Per-clip meta-action labels (high-level motion labels like yield, lane change, stop) are derived directly from trajectory data — no manual annotation of driving decisions. * Keyframe selection at ego action transitions gives the VLM precise decision-making context — the autolabeler runs only on the frames that actually matter, not the entire clip. * Chain-of-Causation reasoning labels via VLM inference produce the training signal that reasoning-based AV models (such as Alpamayo 2 Super) consume — at large scale, in YAML mapping keyframe timestamps to ego behavior descriptions. * Supports both local models and hosted VLM APIs (such as GPT-5, GPT-5.5) — pick the deployment path that matches your data-residency and/or system constraints. * CoC reasoning labels generated by this pipeline (verified by manual quality assurance) are shipped in the NVIDIA Physical AI Open Dataset on Hugging Face (1,728 train + 349 val labels in current release) — start training on day one without re-running the labeler. </div> | Alpamayo | Version 2.0 \| Release date 2026/10/21 | | 1792540800 | {"url" => "https://github.com/NVlabs/alpamayo-recipes/", "text" => "More Information"} | Alpamayo Recipes | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/NVlabs/alpamayo-recipes' target='_blank' class='icon icon-page'>GitHub Repository</a></li> </ul> </div> </div> <div data-markdown="1"> NVIDIA Alpamayo Recipes provides workflows for customizing and optimizing autonomous-driving models, spanning supervised fine-tuning, open-loop reinforcement learning, quantization, and distillation recipes. Data preparation and checkpoint conversion utilities help developers use Physical AI AV data across the recipe workflows. ##### Supported hardware: * Minimum GPU: From single-GPU local development to multi-GPU training clusters; configuration depends on dataset size and model variant. * Host OS: Linux (untested on other operating systems) * Other dependencies: PyTorch 2.8+, Transformers 4.57.1+, DeepSpeed 0.17.4+; gated access via Hugging Face; Physical AI AV Dataset on Hugging Face for representative training data ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVlabs/alpamayo-recipes/issues' target='_blank'>Alpamayo Recipes GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/autonomous-vehicles/alpamayo/766' target='_blank'>NVIDIA Developer Forum — Alpamayo</a></li> </ul> ##### Alpamayo Recipes include the following benefits: * Supervised fine-tuning recipes for Alpamayo 1 and 1.5 use Hugging Face Trainer and DeepSpeed to adapt models to custom driving data. * Open-loop RL recipes for Alpamayo 1 and 1.5 use Cosmos-RL and GRPO to post-train policies on recorded driving data. * The Alpamayo 1.5 10B post-training quantization recipe supports FP8 and AutoQuant (NVFP4 + FP8 mixed precision) with NVIDIA Model Optimizer. * Alpamayo 1 and 1.5 to Qwen 2B distillation recipes to support smaller student models. </div> | Alpamayo | Version 2.0 \| Release date 2026/10/21 | | 1791158400 | {"url" => "https://www.nvidia.com/en-us/ai/cosmos/", "text" => "More Information"} | Cosmos Curator | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/nvidia-cosmos/cosmos-curate' target='_blank' class='icon icon-page'>Github Repository</a></li> <li><a href='https://docs.nvidia.com/cosmos-curator-lha/current/' target='_blank' class='icon icon-page'>Cosmos Curator documentation</a></li> </ul> </div> </div> <div data-markdown="1"> Use NVIDIA Cosmos Curator to process, filter, annotate, deduplicate, and organize video datasets for physical AI model development using GPU-accelerated distributed pipelines. ##### Supported hardware: * Minimum GPU: GPU-accelerated processing required; cluster scale depends on dataset volume. NVIDIA H100 multi-node Ray cluster for production-scale curation. * Host OS: Linux; Docker; optional Slurm or NVIDIA Cloud Functions for orchestration. * Other dependencies: Ray-based distributed runtime (Cosmos-Xenna), Python; storage backend for processed assets. ##### **Community:** Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/nvidia-cosmos/cosmos-curate/issues' target='_blank'>Cosmos Curator GitHub Issues</a></li> <li><a href='https://discord.com/channels/827959428476174346/1469128459061301270' target='_blank'>NVIDIA Developer Discord — Cosmos channel</a></li> </ul> ##### **Cosmos Curator includes the following benefits:** * GPU-accelerated video curation pipeline gives you splitting, annotation, filtering, deduplication, and dataset organization at multi-petabyte scale without writing custom data engineering. * AI-powered captioning, classification, and embeddings turn raw video into training-ready data, with checkpointed video and multimodal embedding recipes for resumable processing. * Distributed-by-design via Cosmos-Xenna on Ray empowers you to run on Slurm clusters `slurm_cli`, NVIDIA Cloud Functions `nvcf_cli`, or locally `local_cli` with the same job definitions — switch infrastructure without rewriting pipelines. * Reference Video Pipelines Guide + Reference AV Pipelines Guide cover multi-camera scenarios; minimal example pipelines let you compose and scale custom stages without fighting framework constraints. * Integrated with the wider Cosmos ecosystem — outputs feed Cosmos Dataset Search, Cosmos Reason 2 fine-tuning, and AV Data Factory workflows, so curation effort compounds across downstream uses. <a href='https://nvidianews.nvidia.com/news/nvidia-announces-open-physical-ai-data-factory-to-accelerate-robotics-vision-ai-agents-and-autonomous-vehicle-development' target='_blank'>Read more</a> about Physical AI Data Factory Blueprint </div> | Cosmos | Version 2.8.0 \| Release date 2026/10/05 | | 1785974400 | {"url" => "https://developer.nvidia.com/drive/setup", "text" => "More Information"} | DriveOS [DRIVE AGX Thor Latest] | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://developer.nvidia.com/docs/drive/drive-os/7.2.5/secure/driveos-linux-sdk/NVIDIA_DriveOS_7.2.5.0_Linux_Early_Access_Release_Notes.pdf' target='_blank' class='icon icon-lock'>Release Notes</a></li> <li><a href='https://developer.nvidia.com/docs/drive/drive-os/7.2.5/public/drive-os-linux-installation-ea/index.html' target='_blank' class='icon icon-page'>Installation Guide</a></li> <li><a href='https://registry.ngc.nvidia.com/orgs/drive/teams/driveos-sdk/containers/drive-agx-linux-nsr-aarch64-sdk-build-x86-earlyaccess' target='_blank' class='icon icon-download'>DriveOS Docker</a></li> <li><a href='https://registry.ngc.nvidia.com/orgs/drive/teams/driveos-sdk/resources/driveos_sdk_thor/files' target='_blank' class='icon icon-download'>Additional Packages</a></li> </ul> </div> </div> <div data-markdown="1"> Install NVIDIA DriveOS 7.2.5 Linux SDK using <a href="https://developer.nvidia.com/drive/docker-containers" target="_blank">NVIDIA DriveOS Docker Containers</a> through NVIDIA GPU Cloud (NGC). This requires Ubuntu 20.04 or later on the host PC. [Documentation](https://developer.nvidia.com/drive/documentation) is available under **DRIVE AGX Thor**. Please review the [DriveOS 7.2.5 Installation Guide for NVIDIA Developer Users](https://developer.nvidia.com/docs/drive/drive-os/7.2.5/public/drive-os-linux-installation-ea/index.html) to finalize your DRIVE AGX System Setup. ##### Supported hardware: * NVIDIA DRIVE AGX Thor Submit questions or feedback in the [DRIVE AGX Thor Forum](https://forums.developer.nvidia.com/c/autonomous-vehicles/drive-agx-thor/drive-agx-thor-general/742). We want to hear from you. <b>DriveOS 7.2.5 includes the following benefits: </b> * A modernized core development stack with Ubuntu 24.04, Linux Kernel Noble 6.8, CUDA 13.3, cuDNN 9.23.0, GCC 13.2, and Yocto 5.0, providing a secure, current foundation for building and optimizing high-performance GPU-accelerated applications. * Improved NVIDIA TensorRT™ 11 inference performance, optimization flexibility, and memory efficiency, helping developers build faster, more capable pipelines for production deep learning models. * TensorRT Edge-LLM enables efficient on-device large language model inference on Thor, with documented quantize, export, build, and infer workflows for bringing speech, text, and AI assistant workloads onto the DriveOS stack. * TensorRT Model Connect, a new platform that provides broad support for AI models on DRIVE AGX hardware \- available on [GitHub](https://github.com/NVIDIA/TensorRT-Model-Connect/tree/main). * Enhanced PVA compute integration provides an updated PVA Algorithms Library, cuPVA-CUDA interoperability, full NvSciSync integration, and real-time utilization reporting, helping developers build low-latency, observable vision processing pipelines on Thor. * GPU Hardware Scheduling and Quality of Service helps developers run concurrent autonomous driving, AI inference, digital cluster, and infotainment workloads within a single VM, while built-in GPU recovery mechanisms help applications continue running without requiring a system reboot. * IDE-integrated Docker workflows and the upgraded Nsight suite give ecosystem developers a more productive environment for building, debugging, profiling, and optimizing DNN workloads. * Added Camera SDK resources to support camera driver development and integration on DriveOS, enabling camera integration developers to accelerate support for custom camera solutions. * A curated set of DriveOS Skills that equip your AI coding agent with step-by-step workflows to set up a DRIVE AGX Thor DevKit and to build, validate, and deploy applications with DriveOS. The Skills are available through [Additional Packages](https://registry.ngc.nvidia.com/orgs/drive/teams/driveos-sdk/resources/driveos_sdk_thor/files) on NGC. * DriveOS 7.2.5. adds software support for DRIVE AGX Thor SKU 10 and SKU 12 Developer Kit hardware revisions with updated Infineon 10 Gb Ethernet PHY and Realtek Ethernet Switch devices, helping ecosystem customers deploy on current Thor boards. See the [Product Compatibility Notification](https://developer.nvidia.com/downloads/drive/docs/drive-agx-thor-devkit-product-compatibility-notification.pdf) for revision details. </div> | DRIVE AGX Thor | Version 7.2.5 \| Release date 2026/08/06 | | 1785801600 | {"url" => "https://github.com/NVlabs/alpamayo-recipes/", "text" => "More Information"} | Alpamayo 2 Super | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://huggingface.co/nvidia/Alpamayo2-Super' target='_blank' class='icon icon-download'>Hugging Face Model</a></li> <li><a href='https://github.com/NVlabs/alpamayo2' target='_blank' class='icon icon-page'>GitHub Repository</a></li> </ul> </div> </div> <div data-markdown="1"> Access NVIDIA Alpamayo 2 Super, an open 34-billion-parameter multi-task foundation model for autonomous vehicle development. It combines the 32-billion-parameter NVIDIA Cosmos 3 Super Reasoner backbone with a 2-billion-parameter diffusion Action Expert, both post-trained with reinforcement learning. A single model reasons across 360-degree surround camera coverage and returns future trajectories, Chain-of-Causation reasoning traces, Meta-Actions, visually grounded scene answers, and reasoning auto-labels. ##### Supported hardware: * Minimum GPU: Single NVIDIA GPU with enough VRAM for the 34B model; released inference path runs in BF16. * Host OS: Linux (untested on other operating systems); CUDA 12.x+. * Other dependencies: Python 3.12 and the uv package manager; gated Hugging Face access to the model and the NVIDIA Physical AI Autonomous Vehicles dataset. Model weights are licensed under OpenMDW-1.1. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVlabs/alpamayo2/issues' target='_blank'>Alpamayo GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/autonomous-vehicles/alpamayo/766' target='_blank'>NVIDIA Developer Forum — Alpamayo</a></li> </ul> ##### Alpamayo 2 Super includes the following benefits: * One foundation model across the entire development workflow handles planning, scene understanding, Meta-Action prediction, model critiquing, and reasoning auto-labeling. * Full-surround perception with temporal memory across frames gives the model consistent driving decisions through merges, intersections, and lane changes. * Reasoning auto-labeling with 2D grounding generates Chain-of-Causation labels on your own fleet data, compressing annotation cycles from months to days. * Designed as a teacher model for distillation into compact models that run on NVIDIA DRIVE AGX Thor inside the vehicle. * Permissive OpenMDW-1.1 licensing covers fine-tuning, derivative models, and commercial redistribution. Distilled models deploy commercially without further permission from NVIDIA. <a href='https://developer.nvidia.com/blog/generate-trajectories-reasoning-traces-and-auto-labels-with-nvidia-alpamayo-2-super/' target='_blank'>Read more about Alpamayo 2 Super</a> </div> | Alpamayo | Version Alpamayo 2 Super \| Release date 2026/08/04 | | 1781568000 | {"url" => "https://github.com/NVlabs/alpamayo-recipes/", "text" => "More Information"} | AlpaGym | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/NVlabs/alpagym' target='_blank' class='icon icon-page'>GitHub Repository</a></li> </ul> </div> </div> <div data-markdown="1"> Use NVIDIA AlpaGym, an open-source reinforcement-learning framework for end-to-end autonomous-vehicle policies. AlpaGym runs a policy in closed-loop inside a simulator, scores the resulting drives, and trains on them — pairing AlpaSim (closed-loop simulator) with Cosmos-RL (distributed rollout and training orchestration), with Alpamayo 1.5 (10B) as the default policy. Early-stage research framework; for production AV stacks, customers can customize AlpaGym for their own infrastructure. ##### Supported hardware: * Minimum GPU: 2× NVIDIA GPUs (default Alpamayo 1.5 10B policy configuration); scales single-node to multi-node via Cosmos-RL distributed orchestration. * Host OS: Linux * Other dependencies: AlpaSim as the closed-loop runtime; Cosmos-RL as the distributed RL backend; Alpamayo 1.5 default policy bundle; UV package manager + Hydra config; Hugging Face + Weights & Biases authentication; Apache 2.0 license. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVlabs/alpagym/issues' target='_blank'>AlpaGym GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/autonomous-vehicles/alpamayo/766' target='_blank'>NVIDIA Developer Forum — Alpamayo</a></li> </ul> ##### AlpaGym includes the following benefits: * Closed-loop RL infrastructure for AV driving policies. Runs a policy in a simulator, scores the resulting behavior, and trains on them. The policy learns from the consequences of its own steering rather than from logged ground truth alone. * Modular default stack: AlpaSim + Cosmos-RL + Alpamayo 1.5. Closed-loop simulator + distributed RL orchestration + 10B policy out of the box. Swap any single component (reward, policy, sim, trainer) without rewriting the others. * Scales from single-GPU to distributed multi-node. Default 10B Alpamayo configuration runs on 2 GPUs; Cosmos-RL distributed orchestration scales the same job out across many nodes. <a href='https://developer.nvidia.com/blog/how-to-post-train-autonomous-vehicle-models-in-closed-loop-with-nvidia-alpamayo' target='_blank'>Read more</a> about AlpaGym </div> | Alpamayo | Version 1.0 \| Release date 2026/06/16 | | 1780531200 | {"url" => "https://github.com/NVlabs/alpamayo-recipes/", "text" => "More Information"} | AlpaSim | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/NVlabs/alpasim' target='_blank' class='icon icon-page'>GitHub Repository</a></li> </ul> </div> </div> <div data-markdown="1"> Use NVIDIA AlpaSim, an open-source autonomous vehicle simulation platform, for development and testing of end-to-end AV policies with realistic sensor modeling via NuRec neural-rendering integration and CatK ML-based traffic agent behavior. ##### Supported hardware: * Minimum GPU: NVIDIA GPU-capable simulation host; throughput scales with GPU count for NuRec/video-model rendering * Host OS: Linux * Other dependencies: Omniverse NuRec as the rendering backend for photorealistic sensor simulation; CatK ML-based traffic agents for data-driven closed-loop traffic behavior; supports Alpamayo 1, Alpamayo 1.5, VaVAM, and Transfuser policies out of the box ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVlabs/alpasim/issues' target='_blank'>AlpaSim GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/autonomous-vehicles/alpamayo/766' target='_blank'>NVIDIA Developer Forum — Alpamayo</a></li> </ul> ##### AlpaSim includes the following benefits: * Pluggable rendering backends let you choose Omniverse NuRec (high-fidelity reconstructions of real scenes) or built-in video-model rendering (including Cosmos3-Nano-Sim-Auto) without rewriting policy code. * Realistic ML-based traffic behavior via CatK gives you data-driven agent reactions superior to rule-based traffic models — agents respond to the ego policy instead of following scripted paths, producing interactive closed-loop scenarios. * Flexible gRPC APIs empower you to integrate with AlpaSim from any language and orchestrate distributed runs with pipeline parallelism, keeping per-step latency low for RL training. <a href='https://developer.nvidia.com/blog/building-autonomous-vehicles-that-reason-with-nvidia-alpamayo' target='_blank'>Read more</a> about AlpaSim </div> | Alpamayo | Version 2026.5 \| Release date 2026/06/04 | | 1780272000 | {"url" => "https://github.com/NVIDIA/NuRec", "text" => "More Information"} | Harmonizer | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/NVIDIA/harmonizer/' target='_blank' class='icon icon-page'>Github Repository</a></li> <li><a href='https://huggingface.co/nvidia/Harmonizer' target='_blank' class='icon icon-download'>Hugging Face Model</a></li> <li><a href='https://catalog.ngc.nvidia.com/orgs/nim/nvidia/containers/harmonizer/' target='_blank' class='icon icon-download'>NIM</a></li> </ul> </div> </div> <div data-markdown="1"> Access NVIDIA Harmonizer — an online generative enhancement framework that transforms renderings from imperfect reconstructed scenes into temporally consistent, photorealistic outputs. Harmonizer distills a pretrained multi-step diffusion model into a single-step, temporally-conditioned enhancer that runs on a single GPU inside online simulators. ##### Supported hardware: * Minimum GPU: 1× NVIDIA GPU for inference at 1024 resolution * Host OS: Linux with Docker * Other dependencies: Pretrained Harmonizer checkpoint + Cosmos-Predict2-0.6B-Text2Image base model ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVIDIA/harmonizer/issues' target='_blank'>Harmonizer GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/omniverse/platform/nurec/' target='_blank'>NVIDIA Developer Forum — Omniverse NuRec</a></li> </ul> ##### Harmonizer includes the following benefits: * Choose temporal inference for consistency across frames, or faster non-temporal inference for individual images. * Addresses three failure modes simultaneously: appearance harmonization (compositing artifacts), artifact correction (NeRF/3DGS rendering noise), and lighting realism (mismatched insertions). * Handles inserted dynamic objects from different scenes: composited assets pick up the host scene's lighting and shadows automatically in a single pass. * Specialized data-curation pipeline with five complementary data sources backs the training mix; the same recipe is documented in the repo for teams that want to fine-tune on their own ODDs. </div> | Omniverse NuRec | Version Harmonizer-cosmos-0.6B V1 \| Release date 2026/06/01 | | 1776643200 | {"url" => "https://github.com/NVIDIA/NuRec", "text" => "More Information"} | Asset Harvester | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/NVIDIA/asset-harvester' target='_blank' class='icon icon-page'>Github Repository</a></li> <li><a href='https://huggingface.co/nvidia/asset-harvester' target='_blank' class='icon icon-download'>Hugging Face Model</a></li> </ul> </div> </div> <div data-markdown="1"> Access NVIDIA Asset Harvester, an image-to-3D system that converts sparse, in-the-wild object views from driving logs into simulation-ready 3D Gaussian splat assets for Omniverse NuRec and AV simulation pipelines. ##### Supported hardware: * Minimum GPU: NVIDIA GPU with ~16 GB VRAM recommended; `--offload_model_to_cpu` flag reduces requirements further. Driver ≥570 (CUDA 12.8 compatible). * Host OS: Linux preferred; GCC 10–13 (tested with GCC 12.3); Conda environment * Other dependencies: Input format: 512×512 RGB images with foreground masks (or auto-generated masks); integrates with NCore / NuRec for scaled data processing. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVIDIA/asset-harvester/issues' target='_blank'>Asset Harvester GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/omniverse/platform/nurec/' target='_blank'>NVIDIA Developer Forum — Omniverse NuRec</a></li> </ul> ##### Asset Harvester includes the following benefits: * Automated 3D object extraction from driving logs gives you reusable assets (vehicles, pedestrians, cyclists, road objects) extracted from sparse video views — synthetic data without manual 3D modeling. * Robustness to heavy occlusion, noisy calibration, and extreme viewpoint bias empowers you to mine assets from in-the-wild driving data that traditional photogrammetry rejects outright. * Built-in Mask2Former segmentation generates per-object masks automatically; you start from raw images, not pre-annotated data, and the pipeline handles the rest. * Multi-stage pipeline (Mask2Former → C-RADIO → SparseViewDiT → Object TokenGS) outputs full 3DGS attributes (position, spherical harmonics, opacity, scale, rotation) ready for downstream rendering. * Drops directly into NuRec pipelines for placement, lighting normalization via Harmonizer, and closed-loop simulation — assets compound across the simulation stack. </div> | Omniverse NuRec | Version 1.0 GA \| Release date 2026/04/20 | | 1773878400 | {"url" => "https://github.com/NVlabs/alpamayo-recipes/", "text" => "More Information"} | Alpamayo 1.5 Nano | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://huggingface.co/nvidia/Alpamayo-1.5-10B' target='_blank' class='icon icon-download'>Hugging Face Model</a></li> <li><a href='https://github.com/NVlabs/alpamayo1.5' target='_blank' class='icon icon-page'>GitHub Repository</a></li> <li><a href='https://catalog.ngc.nvidia.com/orgs/nim/nvidia/models/alpamayo-1-5-10b/-' target='_blank' class='icon icon-page'>NIM</a></li> </ul> </div> </div> <div data-markdown="1"> Access NVIDIA Alpamayo 1.5 Nano, an open reasoning VLA model for autonomous driving that adds expanded input conditioning, text-guided planning, and updated reasoning workflows. The release is available through GitHub and Hugging Face. ##### Supported hardware: * Minimum GPU: NVIDIA GPU with ≥ 24 GB VRAM * Host OS: Linux (untested on other operating systems) * Other dependencies: PyTorch 2.8+, Transformers 4.57.1+, DeepSpeed 0.17.4+; gated access via Hugging Face; model weights under OpenMDW-1.1. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVlabs/alpamayo1.5/issues' target='_blank'>Alpamayo GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/autonomous-vehicles/alpamayo/766' target='_blank'>NVIDIA Developer Forum — Alpamayo</a></li> </ul> ##### Alpamayo 1.5 Nano includes the following benefits: * Navigation conditioning lets you steer the model with explicit navigation inputs, eliminating hard-coded planning rules and surfacing intent throughout the pipeline. * General VQA gives you visual question-answering at inference time — ask the model what it sees or why it chose an action without standing up a separate explanation pipeline. * Flexible multi-camera support enables a single model to work across different camera counts, removing the cost of maintaining separate variants per sensor configuration. * Published post-training scripts deliver a documented path to fine-tune on your proprietary data and deployment targets, replacing months of reverse-engineering with step-by-step recipes. <a href='https://huggingface.co/blog/drmapavone/nvidia-alpamayo-1-5' target='_blank'>Read more about Alpamayo 1.5 Nano</a> </div> | Alpamayo | Version 1.5 \| Release date 2026/03/19 | | 1764720000 | {"url" => "https://github.com/NVlabs/alpamayo-recipes/", "text" => "More Information"} | Alpamayo 1 Nano | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://huggingface.co/nvidia/Alpamayo-R1-10B' target='_blank' class='icon icon-download'>Hugging Face Model</a></li> <li><a href='https://github.com/NVlabs/alpamayo' target='_blank' class='icon icon-page'>GitHub Repository</a></li> <li><a href='https://research.nvidia.com/publication/2025-10_alpamayo-r1' target='_blank' class='icon icon-page'>Research Paper</a></li> </ul> </div> </div> <div data-markdown="1"> Access NVIDIA Alpamayo 1 Nano, an open reasoning vision-language-action model for autonomous vehicle research. The release provides model weights on Hugging Face and source code, inference examples, and fine-tuning scaffolding on GitHub. ##### Supported hardware: * Minimum GPU: NVIDIA GPU with ≥ 24 GB VRAM * Host OS: Linux (untested on other operating systems); Python 3.12.x. * Other dependencies: PyTorch 2.8+, Transformers 4.57.1+, DeepSpeed 0.17.4+; gated access via Hugging Face; model weights under OpenMDW-1.1. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVlabs/alpamayo/issues' target='_blank'>Alpamayo GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/autonomous-vehicles/alpamayo/766' target='_blank'>NVIDIA Developer Forum — Alpamayo</a></li> </ul> ##### Alpamayo 1 Nano includes the following benefits: * Chain-of-Causation reasoning gives you explainable driving decisions with step-by-step rationale for every action, so you can debug, validate safety, and review with humans-in-the-loop without standing up a separate explanation pipeline. * Diffusion-based trajectory decoder outputs 6.4-second plans (64 waypoints at 10 Hz) in the ego-vehicle frame, giving you long-horizon planning that respects vehicle dynamics rather than one-shot heuristics. * Fixed four-camera multi-view input (front-wide, front-tele, cross-left, cross-right). <a href='https://developer.nvidia.com/blog/building-autonomous-vehicles-that-reason-with-nvidia-alpamayo' target='_blank'>Read more about Alpamayo 1 Nano</a> </div> | Alpamayo | Version 1.0 \| Release date 2025/12/03 | | 1761955200 | {"url" => "https://github.com/NVIDIA/NuRec", "text" => "More Information"} | Fixer | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/nv-tlabs/Fixer' target='_blank' class='icon icon-page'>Github Repository</a></li> <li><a href='https://huggingface.co/nvidia/Fixer' target='_blank' class='icon icon-download'>Hugging Face Fixer model</a></li> </ul> </div> </div> <div data-markdown="1"> Access NVIDIA Fixer, a single-step diffusion model for improving rendered novel views and reducing reconstruction artifacts in NeRF and 3D Gaussian splatting workflows, including NuRec simulation pipelines. ##### Supported hardware: * Minimum GPU: NVIDIA Ampere-or-newer GPU (A100, H100, H20, L40 tested). * Host OS: Linux preferred. * Other dependencies: PyTorch ≥ 2.0.0, Docker, Python 3.12; compatible with NeRF and 3DGS outputs. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/nv-tlabs/Fixer/issues' target='_blank'>Fixer GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/omniverse/platform/nurec/' target='_blank'>NVIDIA Developer Forum — Omniverse NuRec</a></li> </ul> ##### Fixer includes the following benefits: * Single-step diffusion reduces reconstruction artifacts and restores detail in rendered NeRF and 3DGS views. * Dual offline-and-online modes let you refine outputs during training-time data preparation, then serve the same enhancer as a real-time neural step during rendering. * Post-trainable on customer data empowers you to fine-tune Fixer for novel ODDs (weather variations, geographic regions, unusual lighting) where the base model needs adaptation. * Linear-attention architecture with a deep compression autoencoder runs at ~26.5 ms per 576×1024 image on H100 — fast enough to drop inline into rendering pipelines. </div> | Omniverse NuRec | Version 2 \| Release date 2025/11/01 | | 1756080000 | {"url" => "https://developer.nvidia.com/drive/setup", "text" => "More Information"} | DriveOS [DRIVE AGX Thor] | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://developer.nvidia.com/docs/drive/drive-os/7.0.3/secure/driveos-linux-sdk/NVIDIA_DriveOS_7.0.3.0_Linux_Release_Notes.pdf' target='_blank' class='icon icon-lock'>Release Notes</a></li> <li><a href='https://developer.nvidia.com/docs/drive/drive-os/7.0.3/public/drive-os-linux-installation/index.html' target='_blank' class='icon icon-page'>Installation Guide</a></li> <li><a href='https://registry.ngc.nvidia.com/orgs/drive/teams/driveos-sdk/containers/drive-agx-linux-nsr-aarch64-sdk-build-x86/tags' target='_blank' class='icon icon-download'>DriveOS Docker</a></li> <li><a href='https://registry.ngc.nvidia.com/orgs/drive/teams/driveos-sdk/resources/driveos_sdk_thor/version' target='_blank' class='icon icon-download'>Additional Packages</a></li> </ul> </div> </div> <div data-markdown="1"> Install NVIDIA DriveOS 7.0.3 Linux SDK using <a href="https://developer.nvidia.com/drive/docker-containers" target="_blank">NVIDIA DRIVE OS Docker Containers</a> through NVIDIA GPU Cloud (NGC). This requires Ubuntu 20.04 or later on the host PC. [Documentation](https://developer.nvidia.com/drive/documentation) is available under **DRIVE AGX Thor**. Please review the [DriveOS 7.0.3 Installation Guide for NVIDIA Developer Users](https://developer.nvidia.com/docs/drive/drive-os/7.0.3/public/drive-os-linux-installation/index.html) to finalize your DRIVE AGX System Setup. ##### Supported hardware: * NVIDIA DRIVE AGX Thor Submit questions or feedback in the [DRIVE AGX Thor Forum](https://forums.developer.nvidia.com/c/autonomous-vehicles/drive-agx-thor/drive-agx-thor-general/742). We want to hear from you. DRIVE OS 7.0.3 includes the following benefits: * Recent Ubuntu, Docker, GCC, C++, Linux Kernel, and Yocto versions give you enhanced performance, improved security, expanded hardware support, and streamlined development workflows. * NVIDIA® CUDA® features like context pause/resume, batched memcpy, and tensor maps let you optimize application performance by providing efficient context switching, reduced CPU overhead for data movement, and precise, hardware-aware tensor descriptions for kernel development. * NVIDIA TensorRT™ 10 empowers developers with enhanced performance through dynamic fusions (Myelin), expanded optimization techniques (ModelOpt with advanced quantization like INT4 AWQ), and improved memory efficiency (NVIDIA Blackwell FP4, INT4 Weight-Only Quantization), alongside greater flexibility for complex models via V3 plug-ins and better debugging capabilities. * TensorRT Edge-LLM offers a pure C++ LLM runtime with minimal dependencies and latencies, supporting FP16, FP8, INT4, and FP4 quantization, speculative decoding for faster responses, LoRA for efficient model customization, dynamic batching for optimized throughput, and out-of-the-box support for popular LLMs and VLMs—all available as an open-source GitHub project. * NVIDIA DriveWorks is deeply integrated into DriveOS releases for streamlined development and is highly optimized to fully leverage Thor SoCs. * Tap into robust error handling mechanisms, including Error Propagation Library (EPL) and System Error Handler (SEH), for enhanced protection and early fault detection, while Secure Boot, Platform Security Controller, Public-Key Cryptography, and Functional Safety Island components collaborate for tightly coupled, integrated security within the platform. [Read more](https://developer.nvidia.com/docs/drive/drive-os/7.0.3/public/NVIDIA_DriveOS_7.0.3.0_Features.pdf) about the features in DriveOS 7.0.3. </div> | DRIVE AGX Thor | Version 7.0.3 \| Release date 2025/08/25 | | 1723420800 | {"url" => "https://developer.nvidia.com/drive/setup", "text" => "More Information"} | DRIVE OS [DRIVE AGX Orin Latest] | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://developer.nvidia.com/downloads/drive/secure/drive-sdk/docs/NVIDIA_DRIVE_OS_6.0.10_Linux_Release_Notes.pdf' target='_blank' class='icon icon-lock'>Release Notes</a></li> <li><a href='https://developer.nvidia.com/docs/drive/drive-os/6.0.10/public/drive-os-linux-installation/index.html' target='_blank' class='icon icon-page'>Installation Guide</a></li> <li><a href='https://registry.ngc.nvidia.com/orgs/drive/teams/driveos-sdk/containers/drive-agx-orin-linux-aarch64-sdk-build-x86/tags' target='_blank' class='icon icon-download'>DRIVE OS Docker</a></li> </ul> </div> </div> <div data-markdown="1"> Install NVIDIA DRIVE&reg; OS 6.0.10 Linux SDK using <a href="https://developer.nvidia.com/drive/docker-containers" target="_blank">NVIDIA DRIVE OS Docker Containers</a> through NVIDIA GPU Cloud (NGC). This requires Ubuntu 20.04 or Ubuntu 18.04 on the host PC. [Documentation](https://developer.nvidia.com/drive/documentation) is available under **DRIVE Orin**. Please review the [DRIVE OS 6.0.10 Installation Guide for NVIDIA Developer Users](https://developer.nvidia.com/docs/drive/drive-os/6.0.10/public/drive-os-linux-installation/index.html) to finalize your DRIVE AGX System Setup. ##### Supported hardware: * NVIDIA DRIVE AGX Orin™ Submit questions or feedback in the [DRIVE AGX Orin Forum](https://forums.developer.nvidia.com/c/autonomous-vehicles/drive-agx-orin/520). We want to hear from you! DRIVE OS 6.0.10 includes the following benefits: * Updated Linux kernel version from 5.15.68 to 5.15.122 * Unified framework to log system level events, UART prints, and error codes onto persistent storage * New Video Anonymization feature that can be enabled/disabled using a flag * Additional encode parameters for PII (Personal Identifiable Information) data </div> | DRIVE AGX Orin | Version 6.0.10 \| Release date 2024/08/12 | | 1694476800 | {"url" => "https://developer.nvidia.com/drive/setup", "text" => "More Information"} | DRIVE OS [DRIVE AGX Orin] | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://developer.nvidia.com/downloads/drive/secure/drive-sdk/docs/nvidia_drive_os_6.0.8.1_linux_release_notes.pdf' class='icon icon-lock' target="_blank">Release Notes</a></li> <li><a href='https://developer.nvidia.com/docs/drive/drive-os/6.0.8.1/public/drive-os-linux-installation/index.html' class='icon icon-page' target="_blank">Installation Guide</a></li> <li><a href='https://registry.ngc.nvidia.com/orgs/drive/teams/driveos-sdk/containers/drive-agx-orin-linux-aarch64-sdk-build-x86/tags' class='icon icon-download' target="_blank">DRIVE OS Docker</a></li> </ul> </div> </div> <div data-markdown="1"> Install NVIDIA DRIVE&reg; OS 6.0.8.1 Linux SDK using <a href="https://developer.nvidia.com/drive/docker-containers" target="_blank">NVIDIA DRIVE OS Docker Containers</a> through NVIDIA GPU Cloud (NGC). This requires Ubuntu 20.04 or Ubuntu 18.04 on the host PC. [Documentation](https://developer.nvidia.com/drive/documentation) is available under **DRIVE Orin**. Please review the [DRIVE OS 6.0.8.1 Installation Guide for NVIDIA Developer Users](https://developer.nvidia.com/docs/drive/drive-os/6.0.8.1/public/drive-os-linux-installation/index.html) to finalize your DRIVE AGX System Setup. ##### Supported hardware: * NVIDIA DRIVE AGX Orin Submit questions or feedback in the [DRIVE AGX Orin Forum](https://forums.developer.nvidia.com/c/autonomous-vehicles/drive-agx-orin/520). We want to hear from you! DRIVE OS 6.0.8.1 includes the following benefits: * Improves <a href='https://developer.nvidia.com/docs/drive/drive-os/6.0.8.1/public/drive-os-linux-sdk/common/topics/security_concepts/PKCS_11Interface48.html' target="_blank">PKCS #11</a> security support by providing: * Support for key ownership and isolation for different user space applications. * ECDSA certificate verification, which enables support for Camera Authentication capabilities. * Client-side interface for importing key data under a secure environment. * Streamlined search capability while preventing permanent memory leaks in secure storage during new object creation. * Enables YUV format transformations via API to provide a lower GPU load by offloading the color transformation function to VIC. * Provides reduced complexity of a sensor (e.g., Camera + Lidar/Radar) fusion algorithm by aligning FSYNC with PTP time. * Enables <a href='https://developer.nvidia.com/docs/drive/drive-os/6.0.8.1/public/drive-os-linux-sdk/common/topics/nvsci/ChiptoChipCommunication1.html#topic_wsn_rrn_hsb__section_gbd_5pl_dwb' target="_blank">PCIe link Error Detection</a> and reporting to improve latency of error detection and perform recovery operations. </div> | DRIVE AGX Orin | Version 6.0.8.1 \| Release date 2023/09/12 | | 1678838400 | {"url" => "https://developer.nvidia.com/drive/setup", "text" => "More Information"} | DRIVE OS [DRIVE AGX Orin] | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://developer.nvidia.com/downloads/driveos-606-linux-release-notes-pdf' class='icon icon-lock' target="_blank">Release Notes</a></li> <li><a href='https://developer.nvidia.com/docs/drive/drive-os/6.0.6/public/drive-os-linux-installation/index.html' class='icon icon-page' target="_blank">Installation Guide</a></li> <li><a href='https://developer.nvidia.com/sdkmanager_deb' class='icon icon-download' target="_blank">SDK Manager</a></li> <li><a href='https://registry.ngc.nvidia.com/orgs/drive/teams/driveos-sdk/containers/drive-agx-orin-linux-aarch64-sdk-build-x86/tags' class='icon icon-download' target="_blank">DRIVE OS Docker</a></li> </ul> </div> </div> <div data-markdown="1"> Install NVIDIA DRIVE OS 6.0.6 Linux SDK. Either install with NVIDIA SDK Manager, **or** NVIDIA DRIVE OS Docker Containers through NVIDIA GPU Cloud (NGC). [Documentation](https://developer.nvidia.com/drive/documentation) is available under **DRIVE Orin**. Please review the [DRIVE OS 6.0.6 Installation Guide for NVIDIA Developer Users](https://developer.nvidia.com/docs/drive/drive-os/6.0.6/public/drive-os-linux-installation/index.html) to finalize your DRIVE AGX System Setup. ##### Supported hardware: * NVIDIA DRIVE AGX Orin 1. NVIDIA SDK Manager: * Requires Ubuntu 20.04 on the host PC. * Install the most up-to-date version of [NVIDIA SDK Manager](https://developer.nvidia.com/nvidia-sdk-manager). 2. NVIDIA DRIVE OS Docker Containers: * Requires Ubuntu 20.04 or Ubuntu 18.04 on the host PC. * Learn more about [NVIDIA DRIVE Platform Docker Containers](https://developer.nvidia.com/drive/nvidia-drive-platform-docker-containers). Submit questions or feedback in the [DRIVE AGX Orin Forum](https://forums.developer.nvidia.com/c/autonomous-vehicles/drive-agx-orin/520). We want to hear from you. DRIVE OS 6.0.6 provides the following benefits: * Introduces Compute Graph Framework (CGF), a graph-based application framework, along with System Task Manager (STM), a static, non-preemptive scheduler. CGF and STM elevate DriveWorks to a fully featured automotive middleware solution. * Enhances PKCS #11 Cryptography API support for multiple cryptographic key slots and tokens to increase application cryptographic key capacity. * Provides a recovery partition backup in the event of catastrophic file system corruption. * Adds an experimental persistent logging mechanism for system-level events, UART logging, and error codes. ##### Additional links: * [DRIVE OS 6.0.6 blog post](https://developer.nvidia.com/blog/running-docker-containers-directly-on-nvidia-drive-agx-orin/) </div> | DRIVE AGX Orin | Version 6.0.6 \| Release date 2023/03/15 | | 1669852800 | {"url" => "https://developer.nvidia.com/drive/setup", "text" => "More Information"} | DRIVE OS [DRIVE AGX Orin] | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://developer.nvidia.com/downloads/drive/secure/drive-sdk/docs/nvidia_drive_os_6.0.5.0_linux_release_notes.pdf' class='icon icon-lock' target="_blank">Release Notes</a></li> <li><a href='https://developer.nvidia.com/docs/drive/drive-os/6.0.5/public/drive-os-installation/index.html' class='icon icon-page' target="_blank">Installation Guide</a></li> <li><a href='https://developer.nvidia.com/sdkmanager_deb' class='icon icon-download' target="_blank">SDK Manager</a></li> <li><a href='https://registry.ngc.nvidia.com/orgs/drive/teams/driveos-sdk/containers/drive-agx-orin-linux-aarch64-sdk-build-x86/tags' class='icon icon-download' target="_blank">DRIVE OS Docker</a></li> </ul> </div> </div> <div data-markdown="1"> Install NVIDIA DRIVE OS 6.0.5 Linux SDK. Either install with NVIDIA SDK Manager **or** NVIDIA DRIVE OS Docker Containers through NVIDIA GPU Cloud (NGC). [Documentation](https://developer.nvidia.com/drive/documentation) is available under **DRIVE Orin**. Please review the [DRIVE OS 6.0.5 Installation Guide for NVIDIA Developer Users](https://developer.nvidia.com/docs/drive/drive-os/6.0.5/public/drive-os-installation/index.html) to finalize your DRIVE AGX System Setup. ##### Supported Hardware: * NVIDIA DRIVE AGX Orin 1. NVIDIA SDK Manager: * Requires Ubuntu 20.04 on the host PC. * Install the most up-to-date version of [NVIDIA SDK Manager](https://developer.nvidia.com/nvidia-sdk-manager). 2. NVIDIA DRIVE OS Docker Containers: * Requires Ubuntu 20.04 or Ubuntu 18.04 on the host PC. * Learn more about [NVIDIA DRIVE Platform Docker Containers](https://developer.nvidia.com/drive/nvidia-drive-platform-docker-containers). Submit questions or feedback in the [DRIVE AGX Orin Forum](https://forums.developer.nvidia.com/c/autonomous-vehicles/drive-agx-orin/520). We want to hear from you. DRIVE OS 6.0.5 provides the following benefits: * Exclusive chip identification to prevent ID collisions and control accessibility * The ability to extract software version in individual bootchains to ensure tools and update the process identify the right images to flash to the target hardware * Support for asymmetric bootchains for customized layout and QSPI * Enhanced PKCS #11 support for secure and accurate implementation of industry standards * Support for closed-box DRIVE Orin Functional Safety Island binaries to minimize the cost of developing custom AUTOSAR firmware </div> | DRIVE AGX Orin | Version 6.0.5 \| Release date 2022/12/01 | | 1665532800 | {"url" => "https://developer.nvidia.com/drive/setup", "text" => "More Information"} | DRIVE OS [DRIVE AGX Orin] | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://developer.nvidia.com/downloads/drive/secure/drive-sdk/docs/nvidia_drive_os_6.0.4.0_linux_release_notes.pdf' class='icon icon-lock' target="_blank">Release Notes</a></li> <li><a href='https://developer.nvidia.com/docs/drive/drive-os/archives/6.0.4/common/public/installation/index.html' class='icon icon-download' target="_blank">Installation Guide</a></li> <li><a href='https://developer.nvidia.com/sdkmanager_deb' class='icon icon-page'>SDK Manager</a></li> <li><a href='https://registry.ngc.nvidia.com/orgs/drive/teams/driveos-sdk/containers/drive-agx-orin-linux-aarch64-sdk-build-x86/tags' class='icon icon-download' target="_blank">DRIVE OS Docker</a></li> </ul> </div> </div> <div data-markdown="1"> Install NVIDIA DRIVE OS 6.0.4 Linux SDK (rev. 1) with DriveWorks feature and quality improvements. This is an update to NVIDIA DRIVE OS 6.0.4 (originally published on 2022/08/29). Either install with NVIDIA SDK Manager **or** NVIDIA DRIVE OS Docker Containers through NVIDIA GPU Cloud (NGC). [Documentation](https://developer.nvidia.com/drive/documentation) is available under **DRIVE Orin**. Please review the [DRIVE OS 6.0.4 Installation Guide for NVIDIA Developer Users](https://developer.nvidia.com/docs/drive/drive-os/archives/6.0.4/common/public/installation/index.html) to finalize your DRIVE AGX System Setup. ##### Supported Hardware: * NVIDIA DRIVE AGX Orin 1. NVIDIA SDK Manager: * Requires Ubuntu 20.04 on the host PC. * Install the most up-to-date version of [NVIDIA SDK Manager](https://developer.nvidia.com/nvidia-sdk-manager). 2. NVIDIA DRIVE OS Docker Containers: * Requires Ubuntu 20.04 or Ubuntu 18.04 on the host PC. * Learn more about [NVIDIA DRIVE Platform Docker Containers](https://developer.nvidia.com/drive/nvidia-drive-platform-docker-containers). * Please activate your access to the NGC ‘drive’ organization by clicking through the NGC activation email you received. Submit questions or feedback in the [DRIVE AGX Orin Forum](https://forums.developer.nvidia.com/c/autonomous-vehicles/drive-agx-orin/520). We want to hear from you. ##### Additional links: * [DRIVE OS 6.0.4 blog post](https://developer.nvidia.com/blog/now-available-drive-agx-orin-with-drive-os-6/) </div> | DRIVE AGX Orin | Version 6.0.4 (rev. 1) \| Release date 2022/10/12 | | 1634688000 | {"url" => "https://developer.nvidia.com/blog/nvidia-driveworks-4-0-now-available/", "text" => "More Information"} | DRIVE OS and DriveWorks [DRIVE AGX Xavier Latest] | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://developer.nvidia.com/drive/xavier#section-drive-os-and-driveworks-nvidia-drive-agx-xavier-latest' class='icon icon-download' target="_blank">DRIVE Xavier Downloads</a></li> </ul> </div> </div> <div data-markdown="1"> Install NVIDIA DRIVE OS 5.2.6 Linux SDK and DriveWorks 4.0. For more information, see the <a href="https://developer.nvidia.com/drive/xavier" target="_blank">DRIVE AGX Xavier</a> page. </div> | DRIVE AGX Xavier | Version 5.2.6 \| Release date 2021/10/20 | | 1611187200 | {"url" => "https://developer.nvidia.com/blog/drive-os-driveworks-updates-streamlined-av-software/", "text" => "More Information"} | DRIVE OS and DriveWorks [DRIVE AGX Xavier] | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://developer.nvidia.com/drive/xavier#section-drive-os-and-driveworks-drive-agx-xavier' class='icon icon-download' target="_blank">DRIVE Xavier Downloads</a></li> </ul> </div> </div> <div data-markdown="1"> Install NVIDIA DRIVE® OS 5.2.0 Linux SDK and DriveWorks 3.5. For more information, see the <a href="https://developer.nvidia.com/drive/xavier" target="_blank">DRIVE AGX Xavier</a> page. </div> | DRIVE AGX Xavier | Version 5.2.0 \| Release date 2021/01/21 | | 1573776000 | {"url" => "https://developer.nvidia.com/blog/nvidia-drive-10-0-now-available/", "text" => "More Information"} | DRIVE Software [DRIVE AGX Xavier] | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://developer.nvidia.com/drive/xavier#section-drive-xavier-latest-version-10-0' class='icon icon-download' target="_blank">DRIVE Xavier Downloads</a></li> </ul> </div> </div> <div data-markdown="1"> Install NVIDIA DRIVE Software 10.0 Linux SDK which includes DRIVE OS 5.1.6.1, DriveWorks 2.2, DRIVE IX, and DRIVE AV. For more information, see the <a href="https://developer.nvidia.com/drive/xavier" target="_blank">DRIVE AGX Xavier</a> page. </div> | DRIVE AGX Xavier | Version 10.0 \| Release date 2019/11/15 | | 1559001600 | {"url" => "https://developer.nvidia.com/blog/drive-software-9-0-now-available-for-download/", "text" => "More Information"} | DRIVE Software [DRIVE AGX Xavier] | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://developer.nvidia.com/drive/xavier#section-drive-software-drive-agx-xavier' class='icon icon-download' target="_blank">DRIVE Xavier Downloads</a></li> </ul> </div> </div> <div data-markdown="1"> Install NVIDIA DRIVE Software 9.0 Linux SDK which includes DRIVE OS, DriveWorks, DRIVE IX, and DRIVE AV. For more information, see the <a href="https://developer.nvidia.com/drive/xavier" target="_blank">DRIVE AGX Xavier</a> page. </div> | DRIVE AGX Xavier | Version 9.0 \| Release date 2019/05/28 | | 1538611200 | {"url" => "https://developer.nvidia.com/drive/drive-px2", "text" => "More Information"} | DRIVE OS with DriveWorks [DRIVE PX 2] | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://developer.nvidia.com/nvsdk-manager' class='icon icon-download' target="_blank">SDK Manager</a></li> <li><a href='https://developer.nvidia.com/drive/drive-px2#drive-downloads' class='icon icon-download' target="_blank">DRIVE PX 2 Downloads</a></li> </ul> </div> </div> <div data-markdown="1"> Install NVIDIA DRIVE OS 5.0.10.3 Linux SDK with DriveWorks for NVIDIA DRIVE PX 2. For more information, see the <a href="https://developer.nvidia.com/drive/px2" target="_blank">DRIVE PX 2</a> page. </div> | DRIVE PX 2 | Version 5.0.10.3 \| Release date 2018/10/04 | | 1792540800 | {"url" => "https://www.nvidia.com/en-us/solutions/autonomous-vehicles/alpamayo/", "text" => "More Information"} | Radar Reprojection | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/NVIDIA/radar-reprojection' target='_blank' class='icon icon-page'>GitHub Repository</a></li> </ul> </div> </div> <div data-markdown="1"> Access NVIDIA Radar Reprojection, an open-source library for radar simulation in autonomous-vehicle development. The data-driven approach originates from NVIDIA's NDAS production simulation system and complements the NuRec ecosystem. ##### Supported hardware: * Minimum GPU: None; radar reprojection runs on the CPU. * Host OS: Linux * Other dependencies: Bazel 7.4.1/Bazelisk; Conda with Eigen 3.4; Python 3.11 and physical_ai_av for dataset examples. Code license: Apache 2.0. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVIDIA/radar-reprojection/issues' target='_blank'>Radar Reprojection GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/omniverse/platform/nurec/' target='_blank'>NVIDIA Developer Forum — Omniverse NuRec</a></li> </ul> ##### Radar Reprojection includes the following benefits: * Reproject logged detections to simulated ego and actor poses within recorded viewpoint coverage, updating geometry and radial velocity for closed-loop testing. * Adaptive multi-frame search combines compatible logged views across the scene as the simulated vehicle diverges from its recorded trajectory. * Load PhysicalAI-Autonomous-Vehicles clips through the included adapter, or connect another dataset through the extensible data-adapter interface. * Integrate the C++ library directly or use the gRPC service alongside GPU-based NuRec camera rendering. </div> | | Version 1.0 \| Release date 2026/10/21 | | 1789171200 | {"url" => "https://github.com/NVlabs/alpamayo-recipes", "text" => "More Information"} | Physical AI AV Dataset | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles' target='_blank' class='icon icon-page'>Hugging Face Dataset</a></li> <li><a href='https://github.com/NVlabs/physical_ai_av' target='_blank' class='icon icon-page'>Physical AI AV Developer Kit</a></li> </ul> </div> </div> <div data-markdown="1"> Download the NVIDIA PhysicalAI-Autonomous-Vehicles dataset, a large, geographically diverse multi-sensor AV dataset for end-to-end driving, neural reconstruction, scenario mining, synthetic data generation, supervised fine-tuning, and reinforcement-learning workflows. ##### Supported hardware: * Minimum GPU: NVIDIA GPU for downstream training, fine-tuning, and neural reconstruction workflows. * Other dependencies: Hugging Face account with acceptance of the NVIDIA Autonomous Vehicle Dataset License Agreement before download. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVlabs/physical_ai_av/issues' target='_blank'>Developer Kit GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/autonomous-vehicles/alpamayo/766' target='_blank'>NVIDIA Developer Forum — Alpamayo</a></li> </ul> ##### Physical AI AV Dataset includes the following benefits: * 1,727 hours of global driving data across 25 countries and 2,500+ cities. A production-grade, geographically diverse multi-sensor dataset with cameras, LiDAR, and radar for robust model training. * Seven-camera coverage on every clip (front-wide, front-tele, cross-left, cross-right, rear-left, rear-right, rear-tele), with LiDAR on 298,326 clips and radar on 160,761 clips. * The 26.09 update corrects offline LiDAR-camera calibration for 115 clips and removes unreliable offline calibration for 25 clips. * Curated scenarios with chain-of-causation labels. Reasoning-grounded annotations let you train explainable driving policies that learn the why behind each action, not just trajectory imitation. ##### Datasets derived from Physical AI AV: * <a href='https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles-NuRec' target='_blank'>PhysicalAI-Autonomous-Vehicles-NuRec</a>: Version 26.10 \| Release date 2026/10/29. Neural reconstructions of 1,601 Physical AI AV clips, built from seven-camera recordings and LiDAR and delivered as USDZ scenes for AV training and closed-loop simulation in NuRec, CARLA, and AlpaSim. Includes per-clip comparisons with NuRec 26.04 using image-fidelity and object-level quality metrics. * <a href='https://huggingface.co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles-Foundation4D' target='_blank'>PhysicalAI-Autonomous-Vehicles-Foundation4D</a>: Release date 2026/10/20. Metric depth and voxel-level occupancy, semantic, and velocity labels for 1,253 Physical AI AV clips, supporting validation of AV perception models and autolabeling tools. * <a href='https://huggingface.co/datasets/nvidia/cascade' target='_blank'>CASCADE</a>: Release date 2026/07/16. Human annotations of causal and spatio-temporal relationships in 2,066 Physical AI AV clips for structured scenario mining and driving-reasoning evaluation. Foundation4D and CASCADE are for non-commercial use; see their Hugging Face pages for access terms. <a href='https://developer.nvidia.com/blog/building-autonomous-vehicles-that-reason-with-nvidia-alpamayo' target='_blank'>Read more</a> about the Physical AI AV Dataset </div> | Alpamayo | Version 26.09 \| Release date 2026/09/12 | | 1761609600 | {"url" => "https://github.com/NVIDIA-Omniverse-blueprints/cosmos-dataset-search", "text" => "More Information"} | Cosmos Dataset Search | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/NVIDIA-Omniverse-blueprints/cosmos-dataset-search' target='_blank' class='icon icon-page'>GitHub Repository</a></li> <li><a href='https://build.nvidia.com/nvidia/cosmos-dataset-search/blueprintcard' target='_blank' class='icon icon-page'>CDS Blueprint page</a></li> <li><a href='https://catalog.ngc.nvidia.com/orgs/nvidia/teams/blueprint/containers/cosmos-dataset-search' target='_blank' class='icon icon-download'>NGC container</a></li> <li><a href='https://catalog.ngc.nvidia.com/orgs/nim/teams/nvidia/containers/cosmos-embed1' target='_blank' class='icon icon-page'>Cosmos Embed1 NIM</a></li> <li><a href='https://docs.nvidia.com/cosmos/cds/latest/documentation.html' target='_blank' class='icon icon-page'>CDS documentation</a></li> </ul> </div> </div> <div data-markdown="1"> Deploy NVIDIA Cosmos Dataset Search, a vector search and visual analytics workflow for ingesting, indexing, searching, and curating large multimodal video datasets. ##### Supported hardware: * Minimum GPU: 1x NVIDIA GPU for Cosmos Embed1 embedding pipeline; NVIDIA H100 for production-scale embedding generation * Host OS: Linux; Docker Compose for local deployment; Helm / Kubernetes for scaled deployments. * Other dependencies: Milvus vector database, NVIDIA cuVS, Cosmos Embed1 NIM as the embedding backend. ##### **Community:** Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVIDIA-Omniverse-blueprints/cosmos-dataset-search/issues' target='_blank'>Cosmos Dataset Search GitHub Issues</a></li> <li><a href='https://discord.com/channels/827959428476174346/1469128459061301270' target='_blank'>NVIDIA Developer Discord — Cosmos channel</a></li> </ul> ##### **Cosmos Dataset Search includes the following benefits:** * Semantic search across video datasets lets you query driving datasets — real or synthetic — by scenario characteristics, weather, geography, and edge cases without manual curation or biased sampling. * GPU-accelerated embedding generation via the Cosmos Embed1 NIM paired with Milvus + NVIDIA cuVS gives you semantic vector search at multi-billion-clip scale. * Automated video ingestion pipeline performs frame extraction, embedding, and metadata indexing without manual tagging — point it at storage and let it run. * Turnkey NVIDIA Blueprint with an interactive React UI, REST API, and CDS CLI empowers different roles (data engineer, researcher, PM) to work the same dataset through the interface that suits them. * Reference deployment on AWS EKS via Helm charts gives you a production-validated path; the same charts adapt to on-prem and other clouds without rewriting orchestration. <a href='https://nvidianews.nvidia.com/news/nvidia-announces-open-physical-ai-data-factory-blueprint-to-accelerate-robotics-vision-ai-agents-and-autonomous-vehicle-development' target='_blank'>Read more</a> about Physical AI Data Factory Blueprint </div> | Cosmos | Version 1.0.0 \| Release date 2025/10/28 | | 1792540800 | {"url" => "https://developer.nvidia.com/blog/build-world-models-for-interactive-simulation-with-nvidia-cosmos/?ncid=so-nvsh-763326&es_id=ff0761982c", "text" => "More Information"} | Cosmos 3 | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/nvidia/Cosmos' target='_blank' class='icon icon-page'>GitHub Cosmos 3 Repository</a></li> <li><a href='https://huggingface.co/collections/nvidia/cosmos3' target='_blank' class='icon icon-page'>Hugging Face Cosmos 3 Base Model Collection</a></li> <li><a href='https://huggingface.co/nvidia/Cosmos3-Nano-Sim-Auto' target='_blank' class='icon icon-page'>Hugging Face Cosmos3-Nano-Sim-Auto Model</a></li> <li><a href='https://huggingface.co/nvidia/Cosmos3-Nano-Transfer-Auto' target='_blank' class='icon icon-page'>Hugging Face Cosmos3-Nano-Transfer-Auto Model</a></li> <li><a href='https://catalog.ngc.nvidia.com/orgs/nim/nvidia/containers/cosmos3/-' target='_blank' class='icon icon-page'>Cosmos 3 NGC NIM Catalog</a></li> <li><a href='https://build.nvidia.com/nvidia/cosmos3-nano-reasoner' target='_blank' class='icon icon-page'>Cosmos 3 Nano Reasoner NGC NIM Catalog</a></li> </ul> </div> </div> <div data-markdown="1"> Use NVIDIA Cosmos 3, a family of omnimodal world models for physical AI that brings world understanding and simulation into a unified Mixture-of-Transformers architecture. Cosmos 3 processes and generates text, images, video, audio, and actions, giving developers a foundation for multimodal reasoning, synthetic data generation, and embodied policy learning across autonomous vehicles and robotics. Cosmos 3 provides model options for different development workloads. Developers can use the base models for understanding and generation, or build on post-trained checkpoints for domain-specific tasks, including autonomous driving simulation and synchronized multi-sensor data generation. ##### Supported hardware: * Minimum GPU: NVIDIA GPU; memory and GPU count depend on the selected model, inference mode, and post-training workload. Refer to the corresponding model card for requirements. * Host OS: Linux * Other dependencies: Python + PyTorch; Cosmos Framework for training and inference, with vLLM, vLLM-Omni, Diffusers, or SGLang depending on the workflow. Cosmos3-Nano is released under OpenMDW 1.1; review each checkpoint's license before use. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVIDIA/cosmos/issues' target='_blank'>Cosmos 3 GitHub Issues</a></li> <li><a href='https://huggingface.co/nvidia/Cosmos3-Nano-Sim-Auto/discussions' target='_blank'>Cosmos3-Nano-Sim-Auto Hugging Face Community</a></li> <li><a href='https://huggingface.co/nvidia/Cosmos3-Nano-Transfer-Auto' target='_blank'>Cosmos3-Nano-Transfer-Auto Hugging Face Community</a></li> <li><a href='https://discord.com/channels/827959428476174346/1469128459061301270' target='_blank'>NVIDIA Developer Discord — Cosmos channel</a></li> </ul> ##### Cosmos 3 includes the following benefits: * Multimodal visual-language reasoning for physical environments. Jointly interpret visual observations alongside language to support spatial and temporal understanding, physical reasoning, and action-related tasks. * Synthetic data generation and closed-loop simulation. Use post-trained autonomous driving checkpoints for synchronized multi-view, multi-sensor generation, supporting both data development and driving-policy evaluation. * A pretrained foundation for action models. Cosmos 3 can serve as the base for developing action models like Alpamayo, adapting its physical-world understanding for downstream reasoning, planning, and action generation. * Open development workflows. Access model weights and Cosmos Framework code for inference, integration, and post-training, with scripts that developers can adapt to their own data, compute, and domain-specific tasks. ##### Cosmos 3 Nano Simulation checkpoints and scripts: Built on the frontier Cosmos 3 Nano model, these post-trained checkpoints support synthetic data generation for autonomous driving and closed-loop simulation rollouts for policy evaluation. Developers can reproduce the workflows using open scripts in Cosmos Framework. * Cosmos3-Nano-Transfer-Auto: A bidirectional checkpoint for synchronized multi-view, multi-sensor generation, enabling synthetic data development for autonomous driving. * Cosmos3-Nano-Sim-Auto: An autoregressive checkpoint for interactive driving simulation with multi-view, multi-sensor outputs, enabling closed-loop rollouts for policy evaluation. <a href='https://developer.nvidia.com/blog/build-world-models-for-interactive-simulation-with-nvidia-cosmos/?ncid=so-nvsh-763326&es_id=ff0761982c' target='_blank'>Read more</a> Cosmos 3 </div> | Cosmos | Version Cosmos 3 \| Release date 2026/10/21 | | 1780272000 | {"url" => "https://github.com/NVIDIA/NuRec", "text" => "More Information"} | InstantNuRec | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://github.com/NVIDIA/instant-nurec' target='_blank' class='icon icon-page'>GitHub Repository</a></li> <li><a href='https://huggingface.co/nvidia/instant-nurec' target='_blank' class='icon icon-page'>Hugging Face Model</a></li> </ul> </div> </div> <div data-markdown="1"> Access NVIDIA InstantNuRec, a feed-forward transformer model that reconstructs dynamic outdoor scenes from autonomous-vehicle driving logs into 3D Gaussian representations in a single forward pass. Run InstantNuRec standalone to convert NCore V4 driving logs to real-time renderable 3D Gaussian Splatting (3DGS) in seconds, or use its output as initialization for higher-fidelity NuRec refinement. ##### Supported hardware: * Minimum GPU: Single NVIDIA GPU with CUDA Compute Capability ≥ 8.0 and ≥ 30 GB VRAM for inference (H100, A100, and RTX 5090 tested). * Host OS: Linux x86_64; native Python installation without Docker. * Other dependencies: Python 3.11, uv, and the repository's pinned PyTorch/CUDA packages; optional rendering requires the render extra. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVIDIA/instant-nurec/issues' target='_blank'>InstantNuRec GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/omniverse/platform/nurec/' target='_blank'>NVIDIA Developer Forum — Omniverse NuRec</a></li> </ul> ##### InstantNuRec includes the following benefits: * Feed-forward 3D Gaussian Splatting in a single forward pass. Converts NCore V4 driving logs to real-time renderable 3DGS in seconds, with no lengthy per-scene optimization. * Export static Gaussian scenes and sky cubemaps, with optional dynamic Gaussians and motion for supported dense inference profiles. * Generates 3D-Gaussian PLY files directly consumable in simulation pipelines and viewable in SuperSplat or NuRec's ply_viewer. * Initialize downstream NuRec training with InstantNuRec output to reduce iteration count and compute on the high-fidelity refinement step. * Built on Depth-Anything-V3, STORM, and BTimer. Leverages production-grade foundation models for monocular depth, dynamic-scene understanding, and temporal modeling. </div> | Omniverse NuRec | Version 1.0 \| Release date 2026/06/01 | | 1792540800 | {"url" => "https://github.com/NVlabs/alpamayo-recipes/", "text" => "More Information"} | AlpaJudge | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://huggingface.co/nvidia/alpajudge' target='_blank' class='icon icon-page'>Hugging Face Model</a></li> <li><a href='https://github.com/NVlabs/alpajudge' target='_blank' class='icon icon-page'>GitHub Repository</a></li> </ul> </div> </div> <div data-markdown="1"> Use NVIDIA AlpaJudge, a family of learned, human-calibrated judges for evaluating autonomous-driving reasoning. AlpaJudge-RB returns a text-only alignment probability in [0, 1] from a question, reference answer, and candidate answer. AlpaJudge-RF evaluates a rationale using six synchronized front-wide/front-tele frames and ego meta-action history and future, without a reference rationale. Both checkpoints run through the standalone alpajudge inference package. ##### Supported hardware: * Minimum GPU: CUDA-capable NVIDIA GPU; approximately 5 GB for weights, plus runtime activations. * Host OS: Linux * Other dependencies: Python 3.12; PyTorch 2.8.0; Transformers 4.57.6; uv package manager; authenticated Hugging Face access. Code license: Apache 2.0; checkpoint license: OpenMDW-1.1. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVlabs/alpajudge/issues' target='_blank'>AlpaJudge GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/autonomous-vehicles/alpamayo/766' target='_blank'>NVIDIA Developer Forum — Alpamayo</a></li> </ul> ##### AlpaJudge includes the following benefits: * State-of-the-art calibrated judges of driving reasoning quality. Score reasoning with or without ground-truth reasoning labels, so evaluation is not gated on hand-written reference rationales. * Interpretable rubric probabilities from the reference-free model. AlpaJudge-RF exposes action agreement, action safety, grounding, completeness, and causality, rolled up into action, component, and overall scores. * Runs end-to-end from an example notebook. Evaluate both judges on a PhysicalAI-AV sample and score policy-generated reasoning without building a harness first. * Reusable as a training signal. Use reasoning quality scores to prioritize training examples, or supply them as an RL reward through a custom adapter. <a href='https://developer.nvidia.com/blog/running-av-reasoning-models-within-onboard-compute-and-latency-limits/' target='_blank'>Read more about Alpamayo 2 Edge & AlpaJudge</a> </div> | Alpamayo | Version AlpaJudge \| Release date 2026/10/21 | | 1794268800 | {"url" => "https://github.com/NVlabs/alpamayo-recipes/", "text" => "More Information"} | Alpamayo 2 Edge | <div class="sidebar" data-markdown="1"> <div class="sidebar__content" data-markdown="1"> ##### Downloads: <ul> <li><a href='https://huggingface.co/nvidia/Alpamayo2-Edge' target='_blank' class='icon icon-page'>Hugging Face Model</a></li> <li><a href='https://github.com/NVlabs/alpamayo2-edge' target='_blank' class='icon icon-page'>GitHub Repository</a></li> </ul> </div> </div> <div data-markdown="1"> Access NVIDIA Alpamayo 2 Edge, a compact vision-language-action (VLA) model for fine-tuning and efficient in-vehicle deployment on NVIDIA DRIVE AGX Thor. Built on the 2B Nemotron-based reasoning backbone of NVIDIA Cosmos 3 Edge and paired with a diffusion action expert, Alpamayo 2 Edge generates driving reasoning and future ego trajectories. ##### Supported hardware: * Validated GPU for PyTorch inference: NVIDIA H100 80GB HBM3 (BF16; PyTorch SDPA and FlashAttention 2). * Onboard deployment target: NVIDIA DRIVE AGX Thor; real-time inference uses quantization and TensorRT Edge-LLM optimization. Reference configuration: DRIVE Thor-X with DRIVE OS 7.2.7.0, CUDA 13.4, and TensorRT 11.3.1.9. * Host OS: Linux * Other dependencies: Python 3.12; uv; PyTorch 2.8.0, torchvision 0.23.0, and Transformers 5.17.0. SDPA is the default; FlashAttention 2.8.3 is optional for standard inference and required for torch.compile/CUDA-graph acceleration. Dataset examples use physical-ai-av 0.2.0 and require Hugging Face dataset access. Code license: Apache-2.0; model-weight license: OpenMDW-1.1. ##### Community: Submit questions or feedback. We want to hear from you. <ul> <li><a href='https://github.com/NVlabs/alpamayo2-edge/issues' target='_blank'>Alpamayo GitHub Issues</a></li> <li><a href='https://forums.developer.nvidia.com/c/autonomous-vehicles/alpamayo/766' target='_blank'>NVIDIA Developer Forum — Alpamayo</a></li> </ul> ##### Alpamayo 2 Edge includes the following benefits: * Build on the first Alpamayo model post-trained with closed-loop reinforcement learning in AlpaGym. * Generate reasoning only when needed, reducing token generation and inference compute during routine driving. * Combine multi-resolution surround-view video, ego history, and navigation instructions to generate CoC reasoning and 64 future poses over 6.4 seconds. * Run PyTorch inference on PhysicalAI-AV clips via CLI or Python API, with optional BF16 acceleration using torch.compile and CUDA graphs. <a href='https://developer.nvidia.com/blog/running-av-reasoning-models-within-onboard-compute-and-latency-limits/' target='_blank'>Read more about Alpamayo 2 Edge & AlpaJudge</a> </div> | Alpamayo | Version Alpamayo 2 Edge \| Release date 2026/11/10 | [Download the raw results data (JSON)](https://developer.nvidia.com/search-data/drive_downloads.json) ## Resources ### Hardware Read about NVIDIA DRIVE® hardware including where to purchase the DRIVE AGX Thor Developer Kit. [Learn More ](/drive/agx) ### Documentation Review DRIVE AGX hardware and software documentation. [Learn More](/drive/documentation) ### Ecosystem Learn about our rich ecosystem of vendors spanning authorized distributors of the DevKit, as well as software, hardware, and service vendors. [Learn More ](/drive/ecosystem-hw-sw) ### Support Ask questions and get answers in our developer forum. [Learn More](/drive/support)