The DUT racing team, has developed a new branche called ‘Driverless', or FSD (Formula Student Driverless) for short. The FSD team will extend last year’s electric racing car DUT-18 with an intricate web of sensors, computers, software and actuators that will take control of the car, basically rendering it into a racing robot.
“The electric competition is about improving the last percentages,” says TU Delft team manager Rutger van den Berg. “But in the driverless competition, we’re still working on solving the big things.”
“The biggest thing is probably state estimation,” says MIT team captain Luke Kulik. “Making sure the car knows where it is and how it moves, making all the sensors fast enough and getting all systems talking to each other.”
Video: Team ETH Zurich
To give an impression of driverless racing, they show a video made by the current leader in the driverless competition, ETH Zurich. Their car slowly moves ahead as it traces the yellow and blue traffic cones marking the sides of the track. But once the car is full circle, it ‘knows’ the way and blasts off full speed ahead, and braking forcefully. On-board computers continuously calculate where to brake and how long to accelerate. The riskiness parameter, or the Max Verstappen factor, may be tweaked. Note how the car hits a cone at 0:46.
The Delft team will modify the car (last year's DUT 18) while the MIT team takes care of computer vision, state estimation (where is the car?), and path planning (how to steer through the curves).
‘Getting there first is what counts’
“About 95% of all the research on autonomous driving is on road behaviour,” says Kulik. “In the FSD competition we don’t bother with other vehicles or hard to predict pedestrians. Getting there first is what counts. That’s what pushes technology forward.”
The race includes a 75 metre acceleration, a small figure of eight track, and a longer race. The longer race starts with an isolated lap over an unknown track, followed by 10 continuous autonomous laps over the same track, called the track drive.
Perception, state estimation, path planning and car control are all software-controlled. Counting on their fingers, the captains estimate that the car has 15 computers and processing units. Three of them have been added to replace the driver.
Anyone who has written any lines of code will appreciate that this kind of software development requires large amounts of testing. This will be done in phases. The first tests are performed on a simulated racing car. The next level is a model car. At the end of February, the teams expect to run the first full-scale tests on the DUT18.
“It will drive extremely slowly at first, just a few metres per second,” says Van den Berg. Preparation for the races in August consists mainly of software improvement and acceleration. Localisation of the cones should not only be accurate, but should also work from a speeding platform. Delays in processing all the sensor data are inevitable, yet they need to be minimised as the vehicle speed increases.
“Up to now, we have been adding all the blocks,” Kulik explains. “From now on, we’ll be optimising the blocks to go as fast as possible.” Most of the software acceleration and improvement will be done in an online collaboration between Delft and Cambridge, Massachusetts.
Although Formula Student is a student-run competition, there are helpful experts on the sidelines. At TU Delft these include Julian Kooij (robotics, computer vision, state estimation); Javier Alonso-Mora (path planning); Bruno Brito (predictive control algorithm); and Riender Happee (vehicle engineering). At MIT the students can fall back on Sertac Karaman, Luca Carlone and Max Opgenoord (indeed, a Dutch TU Delft alumnus).
- The Formula Student team is still hiring. Interested in joining this opportunity in the robotic racing competition? Visit the Interest Drink on Thursday 31 January in the Dreamhall, starting 19:00.