On the streets of San Francisco - Mercedes repositions itself in automated driving

Mercedes is under pressure. Also with autonomous driving. While the Swabians, with the Mercedes Drive Pilot 95 as the first manufacturer in Germany, have brought a Level-3 system onto the streets and thus achieved a veritable coup in public recognition. But in the US and China, the competition is already tinkering with robo-cars at Level 4—and with visible successes. In cities like San Francisco or Austin, Waymo already operates a fleet of commercial driverless taxis that transport passengers around the clock and collect huge amounts of data. In China, beside automakers like BYD, tech companies like Huawei also have this next level of autonomous driving in their sights.

Tesla eyes Europe

The announcements from the Middle Kingdom are not fantasies, as our test drives with Huawei's driver-assistance system show. Even more "frightening" is what BYD's God’s Eye base version delivered in the rain. The autopilot weaved its way not only under the watchful eyes of a driver who regularly had to touch the wheel through the traffic of a highway, but also autonomously steered the car to the exit up to the toll station. And now Tesla is also pushing into Europe with its camera-based FSD system ("Full Self-Driving"). The Americans are already testing robo-taxis in Austin without safety drivers or other occupants. However, not yet as a regular commercial service. The next step should lead the Americans into the lions' den: among others to Germany, France, Italy and Belgium. Still, always with a safety driver at the wheel. Elon Musk has already found a door opener for Europe. Namely the RDW procedure (Rijksdienst voor het Wegverkeer) of the Dutch licensing authority.

The greatest praise for Tesla's FSD came from a professional. Nvidia's Director of Robotics Jim Fan wrote after a drive with FSD v14 ("Full Self-Driving"), that he had at times barely been able to tell whether a human or the neural network, i.e., the machine, was driving. Interestingly, Mercedes cooperates in the United States with Nvidia on its Drive Assist Pro system. In China, however, the German automaker works with the local AI/software firm Momenta. But the regulatory and political climate made the partner switch necessary. Moreover, this could also facilitate deployment in Europe. However, Mercedes pursues a different approach than the competition. Instead of betting on autonomous driving at Level 4, the Stuttgart-based carmaker takes a conservative and seemingly safe path.

Autonomous assistance in hectic city traffic

Don't go all in immediately, but establish a robust helper that lends a hand to people even in hectic city traffic. The car brakes before speed bumps or blinks when it autonomously exits from the roundabout, but the human remains in command at all times. In one year, the system is expected to be production-ready, explicitly defined as Level 2++. That means the driver remains responsible. Behind it lies the concept of "cooperative steering." The human can override the machine at any time. The system recognizes this and learns from the driver's behavior. It remains active even after the intervention. Only when braking does it have to be restarted afterwards. "It's more difficult to develop such a system than a Level-4 system," explains Ali Kani, Vice President Automotive at Nvidia.

The classic model is mathematically well defined. Essentially a clearly defined process chain runs. For example, when the sensors report an obstacle or a red light, the autopilot reacts and initiates the evasive maneuver or brakes. The problem, however, is that even a false alarm is initially forwarded. Moreover, this approach sooner or later reaches its limits. For instance, when remnants of yellow road markings from construction sites are on the asphalt. The end-to-end approach in autonomous driving is data-driven, uses AI, and acts based on the learned model. At Nvidia and Mercedes these two models run in parallel: The end-to-end model is the idea generator for driving, the classical model monitors and checks whether everything fits. The car learns with every meter. Even if a human intervenes. Delicate situations are recorded by the car and the camera images along with the sensor data are analyzed by humans like a video referee. The Nvidia programs then simulate from every incident several variations of the same, which are then fed back into the car.

Computing Power Is Decisive

To accomplish this task, immense computing power is required. Accordingly, the hardware of the rolling computer is configured. Nvidia's main processor can deliver up to 254 TOPS, i.e., trillions of operations per second, and is thus relevant for AI computing power. To ensure that this power does not fade, a fast data network with a transmission rate of 2.5 Gbit/s is necessary. The computer is complemented by more than 600 GB of flash memory. Ultimately, the Mercedes CLA, which is to fulfill these tasks, is the much-touted software-defined vehicle. This is also evident in the sensor equipment of the Mercedes CLA. Among other things: a telephoto and wide-angle front camera for the “double view” forward (wide-angle for the near field or the overview) and the tele lens for more distant objects. In addition four surround cameras for the 360-degree view, as well as four side cameras that monitor cross-traffic and the blind spot during lane changes. What is missing is an expensive LiDAR sensor. Similar to Tesla's case.

Enough of the preface. Now we want to see how the Mercedes-Nvidia-CLA performs in San Francisco's evening rush hour. It quickly becomes clear: The autopilot in the CLA is not a daredevil, but rather a proper actor, freshly out of driving school and adhering to the rules. That is also a good thing; a robo-car in a kamikaze mode would not be particularly trustworthy. We quickly feel comfortable. The machine-driver recognizes that another vehicle is pulling out of a parking space, and politely yields. The Mercedes blinks to change lanes, lets others pass, and then moves over. The fact that the machine has learned how things work in the urban jungle is evident when cars are parked along the right-hand curb. Then the Mercedes proactively shifts to the left, as it is typically faster there.