Motorsports have a long track record of influencing road car development, from the 1912 French Grand Prix-winning Peugeot that was the first car equipped with dual overhead camshafts, to the 3D-printed components that helped Volkswagen set a new record time on the fabled Goodwood hill climb in July.
This year, the Volkswagen Data:Lab attempted to add another motorsport innovation to the list by training a neural network to drive a racecar.
The Munich-based unit was launched by the Volkswagen Group in 2013 to research how data can support each business unit. Most of its experiments remain under wraps, but the lab showed off its autonomous racing efforts at the NetApp Data Days event in Bologna, Italy this week.
The team of data scientists trained a neural network to plan the fastest driving path possible for a vehicle on a range of simulations of racetracks. Its course was then refined using reinforcement learning, with constraints added to guarantee its flexibility and safety, such as obstacles to bypass, high curvature turns to avoid, and limits on speed.
“It’s the same way you would take a young driver, let him train around a lot of tracks for a couple of years, and then he’ll learn how to go fast,” Dr Marc Hilbert, team lead at the Volkswagen Data:Lab, told Techworld.
The system was first developed in a simulator in the cloud, then tested with a platform developed by autonomous driving competition Roborace, before it was finally sent to the racetrack for a true test of its skills.
Integrating what’s written on computers with a car proved challenging as what’s on computer screens in the lab becomes different on a real vehicle. The Data:Lab supplemented their skills in machine learning-powered trajectory planning with the Volkswagen Group’s experience in vehicle dynamics, Roborace’s autonomous driving platform and lidars for real-time obstacle detection developed by automotive design firm Italdesig.
The team put their skills to the test at the ZalaZone racetrack in Hungary.
The first trial involved an emergency stop. As the all-electric DevBot 2.0 vehicle approached an inflatable vehicle at close to 100km per hour it had to suddenly brake as close to possible to the obstacle. It ground to a halt 45 centimeters in front of the inflatable.
They then tested the algorithm’s object avoidance manoeuvre. As the car approached the inflatable, it had to detect the obstacle with lidar and radar and bypass it at the highest speed possible before stopping. The vehicle whistled past the obstacle, missing it by inches.
The final test required the car to adapt its racing line when passing the obstacle and then continue to race down the track.
The arrival of rain added another level of difficulty. The slippery surface lengthened the braking distance, creating conditions that made the car’s performance hard to predict. It achieved it nonetheless.
The safety driver who had been sitting in the car during trials admitted that a human would struggle to match the algorithm’s performance.
“When he came out of the car, he told me he could have never done that kind of manoeuvre. He never could have gone so fast towards the obstacle, and at the last second go so close when passing it,” said Hilbert.
From track to road
The algorithm also made some surprising decisions. The team placed a gate in the middle of the racetrack and asked the car to avoid it. They assumed that the shortest route would be to drive straight through it, but it calculated that because another corner was coming, it would find a faster driving path by driving around it.
Hilbert imagines that future neural networks could be trained to drive in the style of racing legends such as Michael Schumacher, but his ultimate aim is to take these algorithms from race cars to the road.
“These kinds of machine learning methods also can help production cars to be safe on the road, improve performance, and have less maintenance problems,” he said.