A team of researchers from the University of Zurich in Switzerland has recently published their work on training drones with AI to fly in drone racing arenas and outperform championship-level human-controlled drones. After 25 rounds of competition, the drones were able to beat humans in 15 rounds.
Elia Kaufmann, the leader of the research project, explained that the training process for the drones involved pattern recognition of barriers and obstacles present in the racing arena from a set of images, as well as learning flight patterns and positions using Reinforcement Learning. These processes were conducted entirely in a simulated racing arena, until the code was capable of learning and fixing the best and fastest movement strategies.
However, this research aims to present an AI that can compete against humans in a more diverse and challenging environment. It still has limitations, such as the need to compete in enclosed spaces with appropriate lighting conditions that the drones were trained with. Additionally, the drones have not been trained to solve other obstacles, such as collisions with opposing drones, strong winds altering their course, or unclear obstacles.