HeyB in Chess, Go or Stratego – in strategy games, AI is unbeatable. Even in computer games such as car racing, adaptive algorithms now easily outperform the best human players. But what if computer programs left the simulation and gaming environments and had to face real physical competition? Here, too, smart algorithms seem to have the advantage. Researchers from the University of Zurich, in collaboration with scientists from chip maker Intel in Munich, have developed an artificial intelligence system that can beat the best pilots in a drone race for the first time.
In these, human pilots usually guide four-rotor robots, called quadcopters, through a three-dimensional obstacle course. Each participant sees the surroundings of their drone filmed by an onboard camera on their 3D video goggles. It can reach a top speed of 100 kilometers per hour and perform amazing flight maneuvers.
Drone racing is a huge challenge for AI, because drones that fly autonomously are constantly exposed to unexpected situations. Meanwhile, the on-board computer has to process countless data about speed, position and position in space so it can keep the drone on an optimal course.
Until recently, autonomous drones took twice as long to fly through an obstacle course than human-piloted quadcopters, unless they relied on an external positioning system to precisely control their trajectory. But that has now changed with the artificial intelligence system “Swift” developed by Ilya Kaufman and his colleagues.
Aerobatics at speeds of 100 km/h
SWIFT interacts in real time with image data from the on-board camera and current velocity and acceleration data provided by the built-in inertial measurement unit. An artificial neural network uses camera data to determine the drone’s location in space and to recognize gates along the race track. This information is sent to an intelligent controller, which chooses the optimal flight maneuvers for the drone in order to fly across the racetrack as quickly as possible.
However, before the actual race, a certain amount of training time was necessary. SWIFT was trained for an hour on a computer in a simulated environment, where the system taught itself, practicing and improving collision-free flight using trial and error. The system was fed real flight data.
After that, Swift was ready to handle human pilots. The contenders were 2019 Drone Racing League Champion Alex Vanover, 2019 MultiGP Drone Racing Champion Thomas Pitmata and three-time Swiss Drone Racing Champion Marvin Schepper. The competitions took place from June 5 to 13, 2022 on a specially designed racetrack in a hangar at Dubendorf Airport near Zurich. There were seven square targets measuring 25 by 25 metres, which had to be completed in the correct order each round, and in the shortest possible time.
The drones had to fly 75 meters per lap. The three pilots had a week to train. Then everyone competed individually against Swift.
Rule: The first drone to fly three full laps correctly wins the race. In each round, all gates must be passed in the correct order
The final result of the race that the researchers came up with for Kaufman Presented in Nature magazine.SWIFT won 15 of the 25 races and set the fastest lap in the race, half a second short of the best time for a human pilot. The causes of the ten defeats were collisions with the opponent’s drone or with a target. And in 20% of the cases, the autonomous drone was too slow.
Humans are more adaptable
In general, human-controlled quadcopters have proven to be more adaptable than autonomously flying drones. This often fails when the racing conditions are different from those in the training simulation, for example when it is too bright in the showroom.
in Companion commentary Dutch AI researcher and bioelectronic scientist Guido de Kroon of TU-Delft writes that more research in a more realistic and diverse environment will be necessary if one is to exploit the full potential of AI-controlled drone technology. In order to beat human pilots in any racing environment, the system must not be disturbed by external disturbances such as wind, changing lighting conditions, or less obvious obstacles. Researchers too I posted an explanatory video about AI-controlled drones.
For researchers from the University of Zurich, drone racing is not the only application of AI systems such as SWIFT. “If we fly faster, we increase the utility of drones,” says study co-author Davide Scaramosa. This is important, for example, when monitoring forests to cover larger areas in a short time. Fast, autonomous drones can also be used in filmmaking to capture action scenes. “Last but not least, high flight speed can make a critical difference in rescue operations – for example with drones being sent into a burning building.” Last but not least, the military may also be interested in the findings of the Swiss researchers. .
“Total coffee aficionado. Travel buff. Music ninja. Bacon nerd. Beeraholic.”