Swiss researchers invent drone-flying AI that tops champions

Why drone racing isn't like other games?

Previous triumphs of AI systems have included board games like chess and Go or even video games like Gran Turismo. But drone racing is a highly physical sport and follows the laws of physics instead of those coded into simulated environments.

Here, the pilot only has access to first-person views generated from onboard cameras as high-speed drones fly through three-dimensional circuits. Human pilots spend years training themselves to achieve the mastery required to outdo each other in head-to-head contests.

An autonomous drone piloting system needs to constantly keep estimating its speed and location using onboard sensors while flying at high speeds if it needs to beat its human counterparts, and that's exactly what the Swiss researchers have managed to build.

Deep reinforcement learning

Davide Scaramuzza and his team of researchers at the university used deep reinforcement learning to train the autonomous system they call Swift. This approach uses rewards during trial and error in simulation to hasten up the learning process. The AI system was trained using data from a real-world race track designed by a professional drone-racing pilot, and took not more than 50 minutes for the training to be complete.

Once the system was ready, it went head-to-head against three human champions of drone racing, two of whom were world champions of international leagues of the sport. The pilots were also given one week to practice on the race track, after which they went head-to-head against the AI system.

Out of the 25 races held between the system and its human racers, 15 were won by the AI, which also clocked the fastest time on the course, which was half a second faster than the one set by a human.

The test track was laid out in controlled environments, and if the AI had to really beat a human counterpart fair and square, it would also need to take into account various other factors such as wind disturbances, differences in light conditions as well and gates that are usually so not well defined.

The increase in complexity of the real-world scenario would perhaps be overwhelming for the system now, but with further improvements in processing capacity, this cannot be ruled out in the future.

The application of the technology can extend well beyond drone racing and even find use in the military, where drones have been used extensively, experts suggest. Alternatively, the technology can also help power self-driving in vehicles and aircraft as well.

The research findings were published today (August 30) in the journal Nature and can be found here.

Abstract

First-person view (FPV) drone racing is a televised sport in which professional competitors pilot high-speed aircraft through a 3D circuit. Each pilot sees the environment from the perspective of their drone by means of video streamed from an onboard camera. Reaching the level of professional pilots with an autonomous drone is challenging because the robot needs to fy at its physical limits while estimating its speed and location in the circuit exclusively from onboard sensors1. Here we introduce Swift, an autonomous system that can race physical vehicles at the level of the human world champions. The system combines deep reinforcement learning (RL) in simulation with data collected in the physical world. Swift competed against three human champions, including the world champions of two international leagues, in real-world head-to-head races. Swift won several races against each of the human champions and demonstrated the fastest recorded race time. This work represents a milestone for mobile robotics and machine intelligence2, which may inspire the deployment of hybrid learning-based solutions in other physical systems.