In world first, AI beats humans in physical sport

An artificial intelligence (AI) software has defeated three world champions in drone racing, a “new milestone,” according to its developers.

The software, called Swift and developed by researchers at the Zurich University in Switzerland and Intel, won multiple races against its human competitors during tests last year.

While IBM’s Deep Blue won a chess game against Gary Kasparov in 1996 and Google’s AlphaGo bested Go champion Lee Sedol in 2016, this is the first time an AI system has beaten human challengers in a physical sport such as drone racing, according to researchers.

In races that took place in a purpose-built track near Zurich in June of 2022, Swift beat 2019 Drone Racing League champion Alex Vanover, the 2019 MultiGP Drone Racing champion Thomas Bitmatta, and three-times Swiss champion Marvin Schaepper.

According to the paper published in Nature, Swift won several races against each of the human champions and demonstrated the fastest recorded race time – beating the best human performer by half a second.

Until recently, autonomous drones took twice as long as those piloted by humans to fly through a racetrack.

This marks a “new milestone” for AI, researchers said, as physical sports are more challenging for autonomous systems than board or video games because they are less predictable.

“We don’t have a perfect knowledge of the drone and environment models, so the AI needs to learn them by interacting with the physical world,” said Davide Scaramuzza, head of the Robotics and Perception Group at the University of Zurich.

Like its human rivals, the AI system had to control quadcopter drones remotely and fly them at speeds exceeding 100 km/h to win.

The lap was considered finished once racers passed seven square gates in the right order. They also had to complete challenging maneuvers and acrobatic features.

Swift competed based on real-time data collected by an onboard camera, the same as the one used by human racers.

According to researchers, the system trained on a month of simulated flight time, which corresponds to less than an hour on a desktop computer, and taught itself to fly by trial and error.

Training in a simulated environment prevented the destruction of drones in the early stages of learning when the system often crashed, they said.

“To make sure that the consequences of actions in the simulator were as close as possible to the ones in the real world, we designed a method to optimize the simulator with real data,” said first author of the paper, Elia Kaufmann.

According to researchers, human pilots were still better at adapting to changing conditions. AI failed when conditions were different from what it was trained for, even when it only meant there was too much light in the room.

It would matter less when it comes to the technology’s potential real-life applications, where fast-flying and, therefore, battery life-preserving drones would be an advantage in areas such as environmental monitoring or disaster response, researchers said.