Nvidia Announces Mercedes-Benz CLA as the First Vehicle with DRIVE AV Autonomous Driving Technology

[PCauto]Autonomous driving technology is evolving from mere environmental perception to more advanced logical reasoning. The key to this leap lies in NVIDIA's newly launched Alpamayo open-source AI model family.

The first vehicle equipped with this technology is the new CLA model from Mercedes-Benz, a long-time collaborator with NVIDIA. This car will be launched in March 2025 and has attracted market attention due to its impressive range of 800 km.

Alpamayo 1: ChatGPT moment for physical AI

NVIDIA’s Alpamayo open-source AI model has been described as the "ChatGPT moment for physical AI" in the industry. Its primary goal is to address the most challenging long-tail problems in autonomous driving—those rare but high-risk edge cases, such as sudden road obstacles or unconventional behaviors by traffic participants, that typically lead to autonomous driving systems disengaging or failing.

As the flagship model, Alpamayo 1 is a Vision-Language-Action (VLA) model with 10 billion parameters. It breaks away from the limitations of traditional autonomous driving systems, which rely on a three-stage segregated architecture: "perception-planning-action."

Alpamayo 1 employs an end-to-end "chain-of-thought" reasoning model, allowing the system to directly process video streams captured by cameras as input. Using deep learning algorithms, it simultaneously handles the entire process: "environment recognition" (e.g., pedestrians, vehicles, traffic signs), "scene analysis" (e.g., determining traffic congestion or intersections), "logical reasoning" (e.g., 'pedestrian crossing ahead, need to slow down and give way'), and "trajectory generation."

More importantly, Alpamayo 1 synchronously outputs the logical chain behind its decisions, essentially providing an instruction manual for the thought process of the autonomous driving system. This transparency makes it possible for outsiders to understand and trace "why it drives this way."

Alpamayo will be open-sourced

To accelerate the adoption of this core technology, NVIDIA has opted for an open-source approach to lower industry barriers and create an ecosystem advantage.

NVIDIA has freely released the Alpamayo 1 model on the Hugging Face platform. Developers can now fine-tune the model to adapt it to different vehicle types, rather than starting from scratch.

In addition, NVIDIA has introduced the open-source AlpaSim end-to-end simulation framework and the Physical AI Open Datasets, which include over 1,700 hours of complex scene data.

The core value of this simulation framework is its ability to address the high costs of real road testing and the difficulty in replicating edge scenarios. It uses digital twin technology to create virtual road environments that align perfectly with real-world physical laws. Developers can simulate extreme scenarios—such as heavy rain, snowstorms, or sudden emergencies—an unlimited number of times to quickly validate model reliability.

Mercedes-Benz CLA: First Vehicle with NVIDIA DRIVE AV Autonomous System

According to the announced plan, the new Mercedes-Benz CLA, equipped with NVIDIA's full AV system (including Alpamayo inference capabilities), will launch in the U.S. market in the first quarter of 2026, enter the European market in the second quarter, and soon follow in the Asian market.

This system is positioned by Mercedes-Benz as a "Level 2+" driver assistance system, similar to Tesla’s Full Self-Driving (FSD). While it requires continuous driver attention, its long-term goal is to achieve Level 4 autonomous driving capabilities.

Mercedes-Benz has named the system MB.DRIVE ASSIST PRO, which provides full-process city road navigation assistance, from parking lots to destinations. It supports driver intervention at any time through a collaborative steering function. The sensor suite includes 30 components, including 10 cameras, 5 radar sensors, and 12 ultrasonic sensors, providing the hardware foundation for precise perception.

NVIDIA DRIVE AV’s Dual-Stack Architecture for Safety

As part of NVIDIA DRIVE AV's dual-stack architecture, the Alpamayo model is the AI-driven stack responsible for core driving tasks. It handles routine driving decisions and intelligent responses to complex scenarios.

The parallel safety stack, built on the NVIDIA Halos safety system, operates through preset physical rules and safety thresholds, such as automatic warnings when the vehicle exceeds speed limits or forced deceleration when the distance to the car ahead is too close. This forms a basic safety baseline.

In the event of anomalies in the AI-driven stack, the traditional safety stack can immediately take over core safety functions to prevent system failure. This architecture is similar to the dual-engine design in airplanes, ensuring that the failure of a single system does not lead to a complete malfunction.

This architecture not only supports basic driving assistance but also enables advanced features such as point-to-point navigation in complex urban environments and active collision avoidance.

The integrated Orin system-on-chip is a critical computing node, capable of performing 254 trillion calculations per second. It can simultaneously process real-time data from 30 sensors (10 cameras, 5 radars, and 12 ultrasonic sensors), distinguishing between static obstacles and dynamic risk targets, and handling complex scenarios like harsh weather conditions.

PCauto’s Review

Before NVIDIA DRIVE AV, most autonomous driving systems, apart from Tesla’s FSD, relied on predefined rules and fixed response patterns—essentially acting according to a script. When faced with unknown complex situations, these systems lacked flexibility. For example, when an unexpected incident occurs, such as an animal suddenly crossing the road or a vehicle ahead making an illegal lane change, the system often fails and may even disengage completely.

However, NVIDIA DRIVE AV is different. Thanks to the reasoning capabilities of the Alpamayo model combined with the safety assurance of the dual-stack architecture, it breaks the limitations of being program-driven. It does not rely on preset rules. Instead, it "thinks" like a human—first understanding the current road situation and then determining why to drive in a particular way. Even in the face of unfamiliar and complex situations, it can adapt and make flexible judgments.