Recently, we ran a story on the University of Zurich's Robotics and Perception Group, and how its AI drones pretty much smoked human FPV champions in a race. You can read the original story here.
Now, thanks to this article about the host Tattu pilot team of world champs Thomas Bitmatta (2018 MultiGP International Open World Champion),Alex Vanover (2019 DRL World Champion), and Marvin Schäpper (Swiss Drone League Champion), we have an even deeper dive into what happened when humans battled AI at the University of Zurich this past June.
“A Single Lap of 5.112s, Repeated Three Times, Is Beyond Human Reach”
Flying First Person View (FPV) is one of the largest difficulties of operating in a race. The pilot can only get the first view through the FPV glasses and manually operate the flight. Operational stability is low, and the pilot can only see the state of high-speed movement, so a test pilot's ability to adapt to situations is essential.
Above: The AI Drone
In this experimental race, with a week to go before the showdown, the three champion pilots came to the event location in advance, constantly training to familiarize themselves with the course, as well as working with the lab to collect data on eye movements and more. Meanwhile, the test track and drones were specially made for the AI system, as AI Drone has 36 cameras and sensors—all to allow the AI to accurately figure out the best route to get the fastest overall finish time.
“I think on one lap, a human can be faster, but on three or more laps, the AI will win," said Schäpper. "For example, an AI drone can do a 5.112s lap time three times behind each other cause they can just repeat one lap. But a lap time of 5.112s and three repetitions is unachievable by humans. On the track we were using, Thomas got down to 6.26s using the Tattu 4.0 battery, but he managed a 6.066s run using the Tattu R-Line 5.0!"
Above: Tattu 5.0 Battery
The human pilots constantly adjusted their strategies and upgraded their equipment to get faster times.
On June 11, the showdown between AI Drone and the world champions was played out at Swiss Drone Days.
The first round started in the morning, with pilots racing one-on-one with AI Drone in multiple attempts to challenge the fastest lap times.
After several rounds of racing, Vanover's fastest lap time was 6.7s, Bitmatta's was 6.3s and Schäpper scored 5.6s. AI Drone's fastest time in three rounds was only 5.1s. However, during the duel with Vanover, AI Drone had a glitch, hit the gate, and fell down. Vanover had figured out a way to get his drone to stay in the middle of the traversing gate, but the AI Drone's sensors detected an obstacle and couldn't get through. So, the drone flew in the other direction, and caused a crash.
In the race course, the possibility of FPV collisions are high, as the pilots adjust speed and attitude to bypass obstacles and no-fly zones. In AI drones, navigation, guidance, and control are completed through artificial intelligence flight control.
The second round unfolded in the afternoon, which was an image recognition session where both the pilot and AI Drone could adjust the FPV's battery, motor, and propeller for better results.
This time, AI changed the machine, removing the 36 sensing cameras covered in the body and instead installing one camera in the head of the drone, which is equivalent to replacing 36 brains with one brain. The results of this round changed, Vanover and Bitmatta were 5.6s, Schäpper hit 5.44s, and AI Drone, affected by the light recognition, managed only 5.6s. Schäpper finally surpassed AI Drone in this round!
During the competition, the AI constantly adjusted its algorithm by observing the pilot's flight path and eye reaction. Tattu's pilot only beat AI Drone in one score—glory to defeat!
All in all, AI intelligence eventually outperformed humans, but the machine still has limitations, and as technology advances and conditions allow, machine replacement for humans has become possible.