The AI-trained autonomous drone (in blue) achieved the fastest lap overall, half a second ahead of the best time set by a human pilot. Photograph: Leonard Bauersfeld/UZH

AI continues to beat humans in an increasing list of fields. In 1996, IBM’s “Deep Blue” beat Gary Kasparov in chess. In 2016, Google’s AlphaGo beat Lee Sedol, the best Go player at the time. Now, an AI-controlled, 100 percent autonomous drone developed by researchers from the University of Zurich (UZH) has defeated the world champions in drone racing for the first time.

The three world-class champions—2019 Drone Racing League Champion Alex Vanover, 2019 MultiGP Drone Racing Champion Thomas Bitmatta, and three-time Swiss champion Marvin Schaepper—were stunned when the autonomous drone swiftly, as its AI system’s name “Swift” suggests, beat them in first-person-view racing. In this type of racing, human drone pilots control the quadrotors via a headset that is connected to an onboard camera and can reach speeds of over 100 km per hour.

The quadrotors in the race at Zurich were heavily optimized for racing. Each of these drones only weighs 870 g and can accelerate at 4.5 gs, reaching 100 km per hour in less than a second. But this extreme customization also limits their flight duration to less than two minutes. The quadrotor identifies the corners of the racing gates via an Intel RealSense vision system and other visual features to navigate itself on the course. An Nvidia Jetson TX2 module, which includes a GPU, a CPU, and associated hardware, manages all of the image processing and control on board.

The drones start with a 3D map of the venue. Although some might argue that AI-controlled drones have an advantage compared to human drone pilots as they can calculate the optimized trajectory after learning about the map, it is not that easy for AI-controlled drones to perfectly perform the pre-calculated route in real life. “Physical sports are more challenging for AI because they are less predictable than board or video games, explains Davide Scaramuzza, head of the Robotics and Perception Group at the UZH. “We don’t have perfect knowledge about drones and environmental models. The AI ​​has to improve themselves while interacting with the physical world.” The turbulent aerodynamics of a drone flying through a gate and the flexibility of the drone itself all make it difficult to stick to that optimal trajectory.

This is where deep-reinforcement learning, a type of machine learning, comes in. The simulation on which “Swift” was trained helped avoid the destruction of drones in the early stages of the learning process. The real data collected through flying in real life were then fed to the simulator, which in return altered the learning of “Swift” to make it further adapt to real-life situations.

Swift was trained in a simulated environment in which the system taught itself to fly using trial and error. Photograph: Leonard Bauersfeld

After a month of simulated flight time, which equals to less than an hour on a desktop PC, Swift was ready to challenge his human competitors. The track covered an area of ​​25 by 25 meters. There were seven square gates which the drones had to pass through in the correct order to complete a lap. This also included demanding maneuvers such as a split-S, an acrobatic exercise in which the drones half-roll and perform a descending half-loop at full speed.

In the end, “Swift” managed the fastest lap, half a second ahead of the best time set by a human pilot. However, the human pilot proved to be more adaptable than the autonomous drone, which failed when conditions were different than those it was trained on— for example, when the room was too bright.

According to Scaramuzza, drone racing is not the purpose of pushing the boundaries of autonomous flight. “Drones have limited battery capacity; they need most of their energy to stay in the air. If we fly faster, we increase their utility. In applications such as monitoring forests or space exploration, this is important in order to cover large areas in a short time. In the film industry, fast autonomous drones could be used to record action scenes. Last but not least, high flight speed can make a crucial difference in rescue operations—for example, when drones are sent into a burning building.”

• Share