With all its sensor technology, the modified e-Golf* looks a bit like a cute spaceship darting through sunny Hamburg. Also noteworthy: No one has their hands on the wheel. In theory, at any rate – the specially trained driver Wojciech Derendarz attentively monitors the drive and is ready to assume control at any moment. He is sitting in a car that is tasked with providing valuable data about autonomous driving. The various sensors on the roof, on the fenders and in front and back scan the surroundings with lasers, with radar, ultrasound and cameras.
Autonomous Driving Is Pioneering Work
Derendarz is Volkswagen’s project manager for autonomous driving. He was been studying the matter for about ten years. “Everyone in the department for autonomous driving is very excited,” says Derendarz cheerfully. The project is pioneering work, he says. And autonomous driving in the city is the elite discipline in the field. “It’s enormous in terms of diversity and complexity. The key thing is for our artificial intelligence to detect all relevant objects without issuing false alarms,” explains Derendarz. The project in Hamburg has been running since early February. The test route was set up by the city of Hamburg as a 9-km circuit within the city limits. It runs from the Dammtor train station via Messehallen, Landungsbrücken and the Elbphilharmonie to Rödingsmarkt and back.
Learning Process for Team and Machine
But why the field trial? The idea is to test the complete autonomous package on the roads of the Hanseatic city, which is full of traffic lights, turn lanes and traffic signs – not to mention the other traffic participants, who are out in force today: cars, trucks, motorcycle riders, cyclists and pedestrians. What that also means: intersections, right-of-way rules, parking vehicles and lane changes in moving traffic in the space of very short distances. And that is a lot of input for such a computer brain. “Humans have been training their perception and interpretation of sensory input for thousands of generations. Simulating that with a program is a challenge and a learning process for us and the machine,” says Derendarz. The software tasked with mastering the challenge was developed by his team, primarily in the programming languages C++ and Python. Special technology known as GPUs enables parallel programming in which large amounts of input can be processed simultaneously. The project works with different approaches to artificial intelligence: deep learning, neural networks, pattern recognition methods.
How Well Do the Algorithms Do on the Road?
“The combination of various sensors and software modules in the real road environment yields important insights. Some modules are easy to test in the lab or in a simulation. But the key is the interaction of all components,” explains Derendarz. Collecting new data, testing how the algorithms perform on the road, and then readjusting. In this manner, the team is continuously inching its way toward the next levels of autonomous driving. “The transition from the partially automated driving of level two to the conditional automation of level three is a quantum leap,” says Derendarz. After all, the more open a situation is, the more that happens and can happen, the greater the demands on the artificial intelligence that is controlling the vehicle. Parking? The assistant can handle that – it’s a clearly defined task. Keeping within the lane on the highway without rear-ending other vehicles? That’s a shade more challenging, but all vehicles are going in the same direction. And as a rule, there is neither oncoming traffic, nor traffic lights, nor any of the rest of what makes urban traffic so difficult to master.
Safety Is the Top Priority
The e-Golf is modified on the inside as well. The display, for instance, shows stationary or moving objects depending on the situation and settings, as well as lane boundaries and predicted movements of other traffic participants. The system essentially looks ten seconds into the future in order to predict the current situation, be prepared for everything, and even plan its own drive. The situation is re-evaluated based on the data multiple times per second. But the researchers are not ready to let the car drive entirely on its own. In spite of its multitude of sensors, it can still misinterpret things at times. And it is very defensively configured – to make sure that absolutely nothing happens. Safety comes first.
Vehicle Optimally Prepared
It is, nevertheless, astonishing what the car can already do. The car and its computer brain are exceptionally well prepared. The route is precisely measured, with much important data already captured and stored. “A special measuring vehicle has already traveled the route, collecting as much data material as possible. Every lane marking is saved, every curb, every solid line,” explains Derendarz. The system can even communicate with some of the traffic lights via radio connection. “GPS and normal navigation maps would not have sufficed for the purpose – they’re too imprecise. And high buildings of the sort found in urban centers also disrupt reception sometimes.”
People Are Not Always Attentive
Although the Volkswagen test vehicles are already quite capable, it will still be some time before autonomous vehicles for mass use in urban areas come to the market. The next step, as Derendarz sees it, is controlled situations such as autonomous parking in parking garages or automated lane-keeping on the highway. He welcomes the advantages of autonomous driving, which include enhanced safety in addition to time and space savings. “People are not always attentive. Roughly 90 percent of accidents today can be traced back to human error,” says Derendarz. The development could go very quickly, he adds: “In many areas we have already seen exponential leaps in terms of capabilities; sensor technology in particular is proceeding very quickly at the moment. And we are hard at work on the continuing development of autonomous driving.”
*e-Golf: power consumption, kWh / 100 km: combined 14.1 with 17-inch wheels – 13.2 16-inch; combined CO₂ emissions, g/km: 0; efficiency class: A+