Automotive
How automakers and suppliers apply AI and NVIDIA technologies across a vehicle's life: training and validating driver assistance and automated driving, generating rare road scenarios for testing, planning car plants as digital twins and checking quality in production.
The problem
Carmakers work on two hard engineering problems at once. In the vehicle, driver assistance and automated driving software must cope with rare road situations, such as odd obstacles, heavy weather or temporary road works, that test fleets meet too seldom to supply enough training and test data, and each software release needs fresh evidence that it is still safe.
In the plant, every new model or variant has to fit into body, paint and assembly shops that keep building other cars. Clashes between new tooling, conveyors and car bodies were traditionally found in physical trial runs, and checking trim, stitching or part variants by eye is repetitive work in which misses reach customers.
Vehicles are becoming software-defined while program timelines shrink, so design, simulation and validation data must move between the carmaker and many tier suppliers who each own part of the system and its intellectual property.
The approach
NVIDIA describes its automated driving offer as three computers: DGX systems for training, Omniverse and Cosmos running on RTX for simulation and validation, and DRIVE AGX with the safety-certified DriveOS in the car. Cosmos generates long-tail test scenarios and adds weather, lighting and location variations to sensor data, Omniverse NuRec rebuilds recorded drives so they can be replayed in simulation, and Halos is NVIDIA's safety framework spanning vehicle architecture, chips, software and models. Hyperion packages DRIVE AGX compute, DRIVE AV software and a validated sensor suite as a reference architecture.
In production, BMW Group plans its plants in a virtual factory built on Omniverse and says digital twins cover more than 30 production sites, with collision checks for a new model shortened from almost four weeks of physical testing to three days. NVIDIA's BMW case studies describe Omniverse Replicator and Isaac Sim producing synthetic images for visual inspection, and DGX systems training detection models for tasks such as spotting stitching faults and wrong door sill variants. General Motors announced in March 2025 that it will use Omniverse with Cosmos for digital twins of assembly lines and for training factory planning and robotics models; that is a stated plan, not a reported result.
Not every program needs this stack. Many carmakers buy entry-level driver assistance as a complete system from a tier supplier, replay of recorded drives and established scenario tools cover much regression testing, and a single model launch can often be checked in the plant planning software already in use. NVIDIA's training and simulation tools matter most to teams that write their own driving software or launch many models across a plant network.123456
Conceptual architecture
Diagram as a list
Applications & solutions
- Test fleet and customer vehicle sensor dataRecorded camera, radar and lidar drives with vehicle signalsConnects to Data curation and scenario selection
- Closed-loop simulation (Omniverse NuRec, Cosmos scenarios)Replays reconstructed drives and generated rare scenarios against the softwareConnects to Safety case and release evidence
- Virtual factory and line inspection (Omniverse, Isaac Sim)Plans model launches in the plant twin and checks parts on the lineConnects to Model training on DGX systems
Models & frameworks
- Perception, driving and inspection modelsTrained networks that move on to simulation, the car or the plantConnects to Closed-loop simulation (Omniverse NuRec, Cosmos scenarios), In-vehicle computer (DRIVE AGX with DriveOS), Virtual factory and line inspection (Omniverse, Isaac Sim)
Inference & runtime software
- In-vehicle computer (DRIVE AGX with DriveOS)Runs the driving software in real time in the carConnects to Safety case and release evidence
Operations & orchestration
- Data curation and scenario selectionPicks the drives and situations that matter for training and testingConnects to Model training on DGX systems, Closed-loop simulation (Omniverse NuRec, Cosmos scenarios)
- Safety case and release evidenceCollects simulation, track and road results for each software release
Accelerated computing
- Model training on DGX systemsTrains perception, driving and plant inspection modelsConnects to Perception, driving and inspection models
Technologies and their roles
NVIDIA DGX5
Training compute
NVIDIA places DGX as the training computer for automated driving models, and BMW trains inspection and synthetic-data models on DGX systems.
NVIDIA Omniverse3
Driving and factory simulation
Omniverse NuRec rebuilds recorded drives for simulation, and BMW's virtual factory for plant planning is built on Omniverse.
NVIDIA Cosmos6
Scenario and sensor data generation
Cosmos generates long-tail driving scenarios and varied sensor data for training and testing; GM also names it for factory planning models.
NVIDIA Isaac4
Synthetic images for plant inspection
BMW uses Isaac Sim with Omniverse Replicator to create training images, including simulated defects, for a visual inspection system.
NVIDIA AI Enterprise5
Supported AI software in the plant
BMW's DGX case study lists NVIDIA AI Enterprise, with TAO used for training and inference of production vision models.
What you need first1
- Recorded drive data with sensor calibration and timing, collected under clear rules on location and personal data
- A scenario catalog and coverage targets agreed with the safety organization
- Functional safety and safety-of-the-intended-functionality expertise to own the safety case
- 3D scans, CAD and tooling data for each body, paint and assembly shop that will be modeled
- Data-sharing agreements with tier suppliers whose models, parts or software enter simulation
- GPU capacity for training, owned or rented, plus RTX-class machines for simulation
Risks and how to reduce them
- Gaps between simulation and the road
- Use generated and reconstructed scenarios as one source of evidence next to track and public-road testing, and compare their statistics with real fleet data.
- Functional safety of AI components in the vehicle
- Keep a safety architecture with independent monitoring and fallbacks; a safety-certified operating system does not certify the carmaker's own software or the vehicle.
- Personal and location data in fleet recordings
- Blur faces and license plates, minimize stored location history and follow the data protection law of each market where vehicles record.
- Supplier intellectual property in shared models
- Control access to supplier CAD, sensor models and software inside twins and simulations, and set usage rights in development contracts.
- Treating announcements as results6
- Several published automaker programs, such as GM's factory work with NVIDIA, are plans; judge them on later measured outcomes.
Documented examples
BMW Group · Automotive manufacturing
BMW Group: planning car plants in an Omniverse-based Virtual Factory
BMW Group plans and checks its car plants in a Virtual Factory built on NVIDIA Omniverse and OpenUSD. BMW says digital twins now cover more than 30 production sites, cut collision checks for new models from almost four weeks to about three days, and are projected to lower planning costs by up to 30 percent.
Scaling
Related
Sources
- NVIDIA self-driving cars (solutions page) (opens in a new tab)
- NVIDIA Cosmos (product page and FAQ) (opens in a new tab)
- BMW Group scales Virtual Factory (opens in a new tab)
- BMW Group virtual factory (case study) (opens in a new tab)
- BMW optimizes production with AI and DGX systems (case study) (opens in a new tab)
- General Motors and NVIDIA collaborate on AI for vehicles and manufacturing (press release, 18 March 2025) (opens in a new tab)
Thank you. Your correction was sent.
The editors check it against the sources. If you left an email address, they may reply about it.