Front edge: Researchers at the University of Zurich (UZH) have developed a machine learning algorithm for quadcopter control that can outperform professional racing drone pilots. The algorithm calculates “time-optimal trajectories” while taking into account the limitations of the drone.
The feat at first glance seems obvious – machine learning system beat again a man, so what? Nonetheless, professional drone racers are excellent at what they do, and this is the first time that an autonomous system has beaten not one but two world-class human pilots.
To test the system, UZH researchers created a drone flight path (see below). Both the autonomous drone and human pilots were allowed to train on the course. The AI was not only able to show the fastest lap times, but also bypassed two professional pilots by a significant margin at all stages of the route.
AI uses external cameras to track the drone’s trajectory and perform correct calculations. The team hopes to modify the system to use the ATV’s onboard cameras. The use of onboard camera systems is vital for other drone-related tasks. Researchers expect their work to be useful for applications such as search and rescue operations, building inspections, parcel delivery, etc.
The algorithm is also “computationally demanding”. It currently takes a computer up to an hour to accurately calculate the optimal trajectory. Because of this flaw, human pilots are not afraid of replacement, at least for the time being. Clearly, in situations like search and rescue, when time is critical, they will need a program that can calculate their way through waypoints faster.
All technical details are outlined in a team document that was recently published in scientific robotics.
Image Credit: University of Zurich