Interessanter Vergleich der beiden Ansätze von Waymo und Baidu beim autonomen Fahren. Lesenswert!
Keine News mehr verpassen! Jetzt den Newsletter abonnieren .
echnical standards In China, Baidu acts more like a national champion than a private firm – developing technical standards for automated driving together with the authorities. The government coordinates all related activities under the project name “Intelligent Connected Vehicle” (ICV). "I assume we must comply with these ICV standards to gain access to the test regions for automated driving and the infrastructure there," says Volkmar Tanneberger, head of the SAIC-VOLKSWAGEN joint venture, with 25,000 employees in China: "As with electric mobility, foreign companies are gaining relatively late access to the technical specifications, in other words the fundamental framework of automated driving, which they must take into account." This is why leading Western companies are joining the Apollo consortium. In California’s Department of Motor Vehicles’ self-driving car ranking, Google (Waymo) is currently miles ahead. According to 2017 reports, their robot cars can cover distances of over 5,000 miles before a driver has to intervene. But as this progress is now reaching saturation point, the final steps on the road to automated driving will be the most difficult. The question is which technology component will ultimately decide the outcome of the race. Waymo clearly has a head start with its several million test miles. The Californians also started using LiDAR sensors to capture the entire environment around vehicles much earlier than automakers. Unlike cameras, the pulsed laser beams can not only recognize pedestrians by their shape, but also register them with high accuracy using three-dimensional measurement data. Data volume The huge amount of LiDAR data is exactly the problem. In contrast to driver assistance systems, it is not sufficient for autonomous vehicles to detect objects decentrally via respective sensor chips. This was shown by a fatal Tesla accident two years ago, which saw a Model S using Autopilot drive straight into a crossing truck. Following the crash, Tesla issued a statement saying neither the Autopilot nor the driver had noticed the white side of the trailer against a brightly lit sky. "A camera only covers 80 percent of the cases," says Elmar Frickenstein, head of automated driving at BMW: "The remaining 20 percent is where it gets particularly interesting. Different sensors don’t help here. The magic lies in sensor data fusion based on raw data." More than 30 cameras, radars and LiDARS now scan the environment around autonomous test vehicles. They provide gigabytes of data that can be evaluated in real time with the help of image-recognition systems. Only in the last two years have graphics cards suitable for cars had the computing power needed to tackle this amount of information. Mercedes and Bosch use the computing power of six high-end gaming computers to sort the information centrally. For example, neural networks with 128 levels can search each of the 1.3 million measurement points received at a high repetition rate from five LiDAR sensors. The two million pixels of each camera – multiplied by 20 frames per second – are also of great interest to specialists in image recognition systems. Artificial intelligence: different approaches These huge amounts of data cannot be processed without artificial intelligence. Unlike machine learning on the internet, here any mistake could have serious consequences. That’s why there are different strategies in place in the robot car race. Following the Android logic, an "open" platform attracts the most developers, guaranteeing progress by their sheer volume alone. But a car is not a smartphone and such a complex partner network is difficult to orchestrate. Software bugs must not lead to system crashes and if they do, safety-critical systems must be protected by redundancy. As automakers are responsible for the systems, they remain liable in court. Ultimately, they need to master all the technology components themselves – not least because they fear dependence on the IT giants. As with most Apollo partners, BMW has a two-pronged approach. The focus of the collaboration is on the further development of individual technologies, including data analysis and data handling, explains a BMW spokesman. As with the cooperation between Daimler and Baidu, it’s all about networking in the Baidu ecosystem in order to make mobility services as attractive as possible for customers. This includes "voice interaction" between passengers and vehicles and interaction between vehicles and smart homes with the help of devices of interest, BMW adds. Here, a large number of country-specific and licensing requirements had to be clarified. Western car manufacturers in particular are trying to gain deep insights into China without revealing too much of their own know-how. The question remains who will win the race in the end. Baidu will dominate all others in China, that much is foreseeable. In the rest of the world, Google (Waymo) wants to be the big winner – after all, it was the same with Android. The U.S. is still the world leader in artificial intelligence. And for some time now, Sundar Pichai, head of Google’s parent company Alphabet, has been preaching "AI first". Experts estimate that one in two of the world's 100 best AI researchers currently works for Google. On the other hand, the Chinese State Council leaves no doubt that China wants to become an AI superpower. By 2030, the domestic value added is expected to reach one trillion yuan (about 150 billion U.S. dollars). Intelligent cars will play a central role in this national development strategy.