An AI system known as Swift has consistently bested champion drone pilots, including the 2019 Drone Racing League world champion Alex Vanover.
Developed by a team at the University of Zurich, Swift outperformed human pilots 60% of the time, setting new lap records that no human competitor could match.
This technology may pave the way for faster and more efficient drones in various industrial applications.
Drone racing sees high-end quadcopters zoom through obstacle courses at break-neck speeds of over 100km/h.
Pilots utilize headsets to view real-time footage captured by onboard HD cameras as they navigate drones through a series of gates. It’s seriously impressive stuff, and you’d be forgiven for thinking that human instincts are virtually irreplaceable here.
But that’s not the case, as the AI system Swift, which uses deep reinforcement learning (RL), was able to outcompete three top-level human competitors in a series of drone races.
“It managed to beat three top-level human pilots, including Alex Vanover, the 2019 Drone Racing League world champion, 60 per cent of the time,” noted Leonard Bauersfeld, the lead researcher at the University of Zurich.
However, there was one caveat: both human and AI participants were given a week to train on the specific track, leveling the playing field.
Some highlighted that the AI would have no chance if faced with a track without this learning period, so humans still have some edge.
Breaking new ground
Swift’s achievement is significant because it utilizes an entirely on-board computer.
Researchers have attempted to run drones with AI for a decade, but earlier systems required dozens of external cameras and a separate computer to transmit real-time instructions.
The AI system that powers Swift processes video from a single camera to identify gates. It’s super-light-weight and agile, which is ideal.
Classical algorithms then ascertain the drone’s orientation and position based on these gates, and another AI system uses that data to plot the drone’s next moves.
“We fuse the two [computational approaches] in order to get something that is better than just one,” Bauersfeld explains.
AI models have been beating humans in many sports and competitive tasks, such as StarCraft, where AI slammed 99.8% of human gamers in 2019.
AI also scored decisive victories over humans in Chess and Go. IBM’s Deep Blue supercomputer beat Grandmaster Garry Kasparov in 1997, and DeepMind’s AlphaGo beat Go champion Fan Hui in 2015 – though a human player beat AlphaGo earlier this year.
