Car racing with artificial intelligence
Autonomous racing car: Artificial intelligence holds the lane and calculate the best track. Photo: Jacob Kepler/TUM
Driving around the race track at 218 or even 270 kilometres per hour (km/h). What is pure routine for Formula 1 drivers like Lewis Hamilton, Max Verstappen or Sebastian Vettel is a sensation for a racing car that drives autonomously: The car is kept on course merely with the help of sensors and computers and, thanks to artificial intelligence (AI), reacts independently to curves, other vehicles or obstacles even at high speeds: "The critical variable in autonomous vehicles is the latency – the time it takes from the occurrence of an event on the track to the car's reaction," explains Phillip Karle, a research associate at the Department of Automotive Engineering at the Technical University of Munich (TUM) and also team leader of the TUM Autonomous Motorsport team. "In the event of road hazards, a short reaction time makes all the difference and may save lives."
Car manufacturers and logistics service providers have high hopes for autonomous driving vehicles. They are expected to solve transportation, cost and personnel problems one day. The driving functions required for this are regarded as highly complex applications of artificial intelligence and machine learning - and are therefore also welcome teaching and study objects at universities. Motorsport is the appealing vehicle for constructing software for autonomous, fast-responding vehicles and simultaneously competing with R&D teams from all over the world when using them.
For five years, the TUM Chairs of Automotive Engineering and Control Engineering have been jointly developing software for autonomous driving functions. In Team TUM Autonomous Motorsport, around 60 students, PHd candidates, and professors Boris Lohmann and Markus Lienkamp are working on issues such as path planning, environmental perception, motion control, and software and algorithms for evaluating driving and route data. In this way. The software with which modified racing cars were equipped and the international competitions were held was developed step by step. Technical support was also provided by the Leibniz Supercomputing Centre (LRZ): The team used the LRZ Compute Cloud to test the functionality of the vehicle's software architecture and adapt by simulating driving tasks and behavior in many possible scenarios on the race tracks.
Driving commands calculated from data
In doing so, Team Autonomous Motorsport was highly successful. At the end of October 2021, it won first place, as well as one million U.S. dollars in prize money at the Indy Autonomous Challenge in Indianapolis, and at the beginning of January 2022, the team ranked second at the Autonomous Challenge@CES in Las Vegas and 50,000 dollars. "We are super happy with the results, our goal was 200 kilometers per hour, and we achieved that," says team manager Alexander Wischnewski. "During the races, we learned a lot about how parts of the software interact. Research projects often focus on a few specific or isolated issues; here we have the chance to look at the problems of a complete driving system." As in other autonomous vehicles, cameras in Team Autonomous Motorsport's race car, as well as so-called light detection and ranging, or LIDAR, sensors, the electronic global positioning system (GPS) and radar sensors, constantly provide information from the vehicle, track, environment and traffic. A computer records and analyzes this data and, with the help of smart systems and algorithms, uses it to calculate commands for brakes, steering and engine. The team tested the necessary hardware components of their system at the chairs, but a lot of computing time and storage volume is required to work out software and algorithms and to simulate driving situations. To this end, the LRZ provided four computer nodes (CPU) in its cloud, each with 40 cores and a total of 400 gigabytes of data storage. "At night, we ran automated tests on them to detect sources of errors in the basic functions of the software, and we were also able to track their progress and optimize them using test metrics," reports team leader Karle. Team manager Wischnewski adds, "We put a lot of time and energy into simulating the race car and tracks and were able to fix many bugs through virtual races." This also made it easier to implement the software in the real vehicle, and the team also simulated races with up to eight vehicles at the same time, training the autonomous software to react quickly and safely.
Open Source Software and a Start-up
The team was well prepared for those competitions. While in Indianapolis the self-driving cars did their laps alone and the focus was on speed, in Las Vegas two cars competed against each other. This made conditions much more difficult for the on-board system, which has to take into account possible driving errors by the competitor and overtaking maneuvers. Nevertheless, the top speed of 218 from the first race was increased by more than 50 kilometers per hour to 270: "First we drove controlled overtaking maneuvers, then we increased the speed bit by bit," says Karle. "However, when it came to the interaction of perception, motion planning and control, there were minor problems that eventually caused the vehicle to go off track. But it was the first time we competed against another racing vehicle at such high speed, and we deliberately wanted to test limits."
Risk-taking and preparation paid off. The successful placings and practical experience whet the appetite for more: the technology developed should now be able to prove itself in real traffic and transport situations. To this end, Team Autonomous Motorsport has made its research and development work as well as the algorithms freely available on the Internet as open source so that they can be used and further developed by others. Together with colleagues from the team, team manager Wischnewski is also currently founding the start-up Driveblocks. Its business objective is to translate the practical experience gained from research and racing into commercially viable software for autonomous driving. (vs)
Team Autonomous Motorsport of TUM. Photo: J. Kepler/TUM