A drone piloted by artificial intelligence he has beaten multiple human competitors, including the DRL World Champion. An article published in the magazine Nature details how researchers from the University of Zurich and Intel Labs built the first autonomous drone able to beat a person in races 1 on 1.
known as swift, the vehicle uses a combination of algorithms and a perception system that analyzes the environment and assigns control commands. Through deep reinforcement learning, the scientists managed to get the drone to beat three champions of various categories on a track with various obstacles. Swift was able to take down not only her opponents, but also record the best lap times in the competition.
How could he get it? According to researchers from the Robotics and Perception Group at the University of Zurich, Swift uses an inertial sensor, a camera and an on-board computer. The drone uses visual inertial odometry (VIO) to estimate its position, speed, and orientation. With the help of the camera and a PNP algorithm, Swift triangulates the location of the doors she must pass through during each lap.
The drone uses a control policy that takes the data produced by the sensing system and translates it into control commands. Before being matched against a human competitor, the policy is trained using deep reinforcement learning. Inside a simulator, investigators run a sequence with 100 agents in parallel to find the fastest path on the track.
AI drone vs. a world champion who is better?
To test Swift, scientists built a track designed by Alex Vanover, DRL World Champion in 2019. The circuit consists of a 75m lap where the pilots must pass through seven square gates piloting drones similar to those used in competitions.
After a week of practice, Vanover was joined by two-time MultiGP champion Thomas Bitmatta, as well as three-time Swiss national champion Marvin Schaepper. The drivers competed against Swift in three-lap 1-on-1 races where the requirement was to go through the doors in the correct order. After several races, the AI-powered autonomous drone claimed victory and posted the fastest lap times.
Although the results are surprising, Swift has a slight advantage over her human opponents. While Vanover and company only use the camera and their abilities to control the vehicle, the autonomous drone uses an inertial sensor and has a slower response time in sensory and motor reactions. The only downside to Swift is that the camera has a 30Hz refresh rate, while human pilots use a 120Hz refresh rate.
“It feels different to compete against a machine, because you know that it doesn’t get tired”
The results obtained by Swift are impressive and represent an important step in deep reinforcement learning techniques. Previously, artificial intelligence managed to outperform humans in games like go, StarCraftand in recent times, Grand Touring. All of them were limited to a virtual environment, so the jump to a physical space means a great advance.
Another noteworthy detail is that Swift does not require robust hardware for training. The simulator runs on a computer with a Core i9-12900K processor, 32 GB of RAM and a GeForce RTX 3090 GPU. According to the study, it took the scientists 50 minutes before discovering the ideal route in the simulator.
“The possibilities are endless, this is the start of something that could change the whole world. On the other hand, I am a runner, I don’t want anything to be faster than me,” said Thomas Bitmatta. “Soon, the AI drone could even be used as a training tool to understand what would be possible.”