Autonomous Racing Breakthrough
A team from the University of Zürich (UZH) and Intel has recently developed an autonomous drone system, Swift. This breakthrough system outpaces human champions in first-person view (FPV) drone racing. Notably, AI systems have previously surpassed human abilities in intellectual games like chess, Go, and StarCraft. Yet, this marks a significant milestone as an AI system beats humans in a physically-oriented sport.
FPV Drone Racing Explained
FPV drone racing is a competitive activity where participants steer high-speed drones through a designated obstacle course. Pilots utilize remote controls and wear a headset, which offers a video stream from a drone-based camera. This headset renders a direct perspective of the drone’s view, immersing pilots in the race.
Swift: The New AI Racing Sensation
The UZH research group had been keen on developing an optimal AI-driven drone pilot for years. However, a reliance on motion-capture systems hampered previous attempts. Their latest innovation, Swift, relies primarily on machine vision, making the AI’s capabilities more comparable to human pilots.
Swift gathers real-time data through an onboard camera, similar to those utilized by human racers. It includes an integrated inertial measurement unit to track the drone’s acceleration and speed. This data is then processed by an artificial neural network to determine the drone’s spatial position and locate race gates. Subsequently, a deep neural network-based control unit identifies the optimal racing strategy. Swift’s design utilized reinforcement learning in a simulated environment, enabling the system to self-train through trial and error.
Head-to-Head with Champions
Swift faced off against renowned human pilots: Alex Vanover (2019 Drone Racing League champion), Thomas Bitmatta (2019 MultiGP Drone Racing champion), and Marvin Schaepper (three-time Swiss champion). The races spanned from June 5 to June 13, 2022, on a specialized track. This track, a 25-by-25-meter area, featured seven square gates in a specific sequence for lap completion. The AI drone secured several victories, even recording the race’s fastest lap time. However, Swift displayed shortcomings in adaptability, especially under varying light conditions. Thus, while AI has significantly evolved in navigating tangible surroundings, humans still excel in adaptability.
The potential of Swift’s Technology
The research team, comprising of Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio, Matthias Müller, Vladlen Koltun, and Davide Scaramuzza, shared their findings in a Nature article titled “Champion-level drone racing using deep reinforcement learning.” According to them, Swift’s technology promises various practical applications. Flying drones at faster speeds augments their usability, considering their confined battery life. Potential applications span from forest monitoring and space exploration to film industry use, particularly in capturing fast-action sequences. Moreover, this technology could optimize search-and-rescue operations, enabling drones to quickly survey extensive areas.