Fully Autonomous Racing Drone Outperforms Two World-Class Pilots

It takes a lot of skill to pilot a high-speed quadcopter through confined spaces and around sharp bends at great speeds. Up until recently, humans were unrivalled in their drone racing abilities – but that all just changed as scientists at the University of Zurich have developed an AI algorithm that can fly a drone around a racecourse faster than professional drone racing pilots.

The algorithm calculates “time-optimal trajectories” as well as the drone’s limitations. It was developed as part of a push to improve the capabilities of autonomous drones and having an algorithm that is capable of outperforming the best human pilots at high speeds is a huge sign of progress for the researchers.

To test the autonomous drone’s capabilities, researchers from UZH set up a racecourse (shown in the video below) and allowed both the human pilots and the autonomous drone to train on the course beforehand. The AI managed to beat the two professional pilots through every single waypoint by significant margins, which obviously resulted in the fastest lap time.

Watch a video of the autonomous racing drone in action:

The algorithm made use of external cameras to track the drone’s position on the course while making calculations to optimise its path. The UZH team plans to modify the system in future to make use of the quad’s onboard cameras instead which is important if autonomous drones are to be applied to practical uses. The researchers hope to make these autonomous drones useful for search and rescue, package delivery and inspection among other things.

The team still has their work cut out for them though because, in addition to currently relying on external cameras, the algorithm requires about an hour of “computationally demanding” time to calculate the most “time-optimal” trajectory along the track. Once these factors are addressed, we will likely see autonomous drones finding their way into commercial use.

All the technical details have been published in a paper by the team in Science Robotics.

Images from Robotics and Perception Group, UZH