The Swift AI employed a method known as deep reinforcement learning, securing victory in 15 out of 25 races against world champions
After dominating humans in activities ranging from chess and Go to StarCraft and Gran Turismo, artificial intelligence (AI) has elevated its prowess and triumphed over world champions in an actual sport.
The most recent individuals to experience defeat at the hands of AI are three skilled drone racers who were surpassed by an algorithm capable of navigating a drone through a 3D race course at high speeds without frequent crashes, or at least not crashing too often.
Created by researchers at the University of Zurich, the Swift AI emerged victorious in 15 out of 25 races against world champions, securing the fastest lap on a course where drones reach speeds of 50 mph (80 km/h) and endure accelerations up to 5g, potentially causing blackout in many people.
“This outcome represents the inaugural instance of an AI-driven robot surpassing a human champion in a genuine physical sport specifically designed for and by humans,” stated Elia Kaufmann, a researcher involved in the development of Swift.
First-person view drone racing entails maneuvering a drone through a course marked with gates that must be navigated cleanly to prevent a crash. Pilots perceive the course through a video feed from a camera affixed to the drone.
In a publication in Nature, Kaufmann and his team detail a sequence of one-on-one races between Swift and three champion drone racers: Thomas Bitmatta, Marvin Schäpper, and Alex Vanover. Leading up to the competition, the human pilots had a week to practice on the course, whereas Swift underwent training in a simulated environment featuring a virtual duplicate of the course.
Swift utilized a technique known as deep reinforcement learning to determine the optimal commands for navigating the circuit. Due to the trial-and-error nature of this method, the drone experienced hundreds of crashes during training. However, as it was a simulation, researchers could easily restart the process.
In a race, Swift transmits video from the drone’s onboard camera to a neural network that identifies the racing gates. This information, coupled with data from an inertial sensor, is used to estimate the drone’s position, orientation, and speed. These estimates are then processed by a second neural network, determining the commands to be sent to the drone.
Race analysis revealed that Swift consistently exhibited faster starts and executed tighter turns than human pilots. The swiftest lap by Swift clocked in at 17.47 seconds, surpassing the fastest human pilot by half a second. However, Swift was not infallible; it lost 40% of its races against humans and encountered crashes. The drone appeared to be sensitive to environmental changes such as lighting.
The world champions experienced a range of emotions after the races. “This is the start of something that could change the whole world. On the flip side, I’m a racer, I don’t want anything to be faster than me,” expressed Bitmatta. Schäpper noted, “It feels different racing against a machine because you know that the machine doesn’t get tired.”
A noteworthy advancement is that Swift can effectively navigate real-world challenges like aerodynamic turbulence, camera blur, and changes in illumination, issues that can perplex systems attempting to follow a pre-computed trajectory. Kaufmann mentioned that a similar approach could assist drones in searching for people in burning buildings or conducting inspections of large structures, such as ships.
While the military maintains a keen interest in AI-powered drones, there is skepticism regarding the immediate implications for warfare. Dr. Elliot Winter, a senior lecturer in international law at Newcastle Law School, cautioned, “We must be careful not to assume that advancements such as these can easily be transplanted into a military context for use in military drones or autonomous weapons systems, which are involved in critical processes such as target selection.
Alan Winfield, a professor specializing in robot ethics, acknowledged the “inevitable” military applications of AI but expressed uncertainty about how the recent advancements could specifically benefit the military, aside from the potential use of drone flocks closely following a plane.
Kaufmann shared a similar skepticism, stating, “Almost all drones are used in wide-open battlefields and are either used for reconnaissance or as weapons against slow-moving and stationary targets.