NVIDIA is finally letting its AI take the wheel on U.S. public roads. The company announced its DRIVE AV software will debut its “address-to-address” L2++ autonomous driving capabilities in the all-new Mercedes-Benz CLA, expected by the end of this year. In a striking new demo, the system navigated the chaotic streets of San Francisco using only cameras and radar, a feat NVIDIA attributes to its dual-stack AI running on a processor capable of 254 trillion operations per second (TOPS).
In the demonstration, NVIDIA’s VP of Autonomous Driving Software, Sarah Tariq, showed the system handling double-parked cars, unpredictable pedestrians, and unprotected left turns with an unnervingly human-like smoothness. The secret sauce is a dual-stack architecture: a traditional, safety-first classical system runs in parallel with an end-to-end AI model trained on “countless hours of human driving.” The car then picks the safest and most comfortable path, which, according to Tariq, is usually the one chosen by the AI. This entire operation runs on a single NVIDIA DRIVE AGX Orin system-on-a-chip, pointedly skipping expensive LiDAR and perpetually outdated HD maps.
The new CLA, one of the first cars built on the new Mercedes-Benz Operating System (MB.OS), recently earned a five-star EuroNCAP safety rating. It scored an impressive 94% in Adult Occupant Protection and 85% in Safety Assist systems, with the performance of its NVIDIA-powered active safety features contributing to the top score.
Why is this important?
This marks a significant leap toward the “AI-defined vehicle.” By running on the new chip-to-cloud MB.OS platform, the system is designed for over-the-air (OTA) updates, meaning the car’s driving skills could theoretically improve while it’s parked in the garage. NVIDIA isn’t just building a driver-assist feature; it’s creating a continuously learning platform. The company is leveraging its “cloud-to-car” pipeline, using NVIDIA Omniverse to create digital twins of factories and run over a million virtual test replays a day to validate software. This approach, which turns real-world miles into billions of simulated ones, is how NVIDIA plans to scale its tech not just for Mercedes, but for partners like Jaguar Land Rover, Lucid, and Stellantis.
