An autonomous drone, piloted by Artificial Intelligence, surpassed racing champions for the first time, according to a study published in the journal Nature this Wednesday (30).
This feat paves the way for the optimization of systems used in autonomous vehicles, or industrial robots.
The race was held on a 75-meter circuit made up of seven portals that must be crossed in a pre-determined order, with machines that easily reach 100 km/h and accelerations that would leave an F1 behind.
Three champions of the sport were recruited by the Robotics and Perception Group at the University of Zurich to face the drone.
Equipped with helmets that transmitted images from the drone they were piloting, the three men, including a former world drone racing league champion, had a week to prepare.
The autonomous drone won the most races against each of them and completed the fastest lap of the circuit.
This is the first time that “an autonomous mobile robot has achieved world champion-level performance in a real-world competitive sport,” according to the study.
Some drones would have reached an “expert” level, but with the help of an external motion capture system to optimize their trajectory.
This was an “unfair” advantage for the Zurich team that presented Swift, a completely autonomous system that takes only its sensors and calculation power on board the drone.
“Swift corrects its course in real time, sending 100 new orders per second to the drone”, explains doctoral student Elia Kaufmann, main author of the study, to AFP.
Swift’s secret relies on a technique called deep reinforcement learning, which combines processing large amounts of data with observing rules that reward the machine’s progress.
The system has tested millions of trajectories, matching your awareness of your surroundings and your progression to the next portal. “Swift trained the equivalent of a month of real time, but in acceleration, that is, in one hour on a computer”, added Kaufmann.
The machine has some inherent advantages, such as a center that provides it with information, such as acceleration, that the human pilot cannot feel without boarding a drone. Another advantage is the reaction time to an order: five times faster than the response of the human brain.
Humans, on the other hand, have an advantage in a degraded environment, for example, when there are changes in light, something that the Swift would have difficulty managing. Humans also have the advantage of reducing speed and avoiding accidents. The machine, on the contrary, always goes to the maximum, “taking many risks”, affirms the study.
The impact of these works extends beyond drone racing, says Guido de Croon, an expert on the subject and professor at the Technological University of Delft, Netherlands, in a commentary accompanying the study in Nature.
According to him, advances in this field are of great interest to the military, but “there is a much wider range of applications”.
For Elia Kaufmann, who today works as an engineer at a company that manufactures drones for industry, the challenge is to respond to “an inherent weakness of autonomous drones: a very limited flight range”.
The focus adopted with Swift, “which allows replanning actions in real time without the need to recalculate a trajectory”, would thus allow for more efficient navigation and, therefore, energy savings.