The AI-powered drone is the fastest in the world

A drone driven solely by an AI application beat world champions in human-powered drone racing, a result that seemed unattainable. Real-world applications of this fast drone include environmental monitoring or disaster response, among others. The technological feat can be compared to the triumph of the IBM computer Deep Blue over the great chess player Garry Kasparov in 1996.

A team of scientists from the University of Zurich in Switzerland, together with experts from the Intel company, achieved to develop an artificial intelligence (AI) system capable of autonomously controlling a drone and breaking all speed records: In a recent test against the world’s human-powered champion drone, it clearly won, beating the other competitors’ best scores by half a second.

AI is taking a big step forward

The technological feat can be compared to the triumph of IBM’s Deep Blue computer over the great chess player Garry Kasparov in 1996: these are turning points where technology, in this case AI, it seems to go beyond its own limits and become ready for new challenges and applications. And as we’ve seen over the course of this year and last year, progress and development suggest that artificial intelligence has no upper limit.

Now he Swift system he managed to win multiple events against three world class racing champions. drones with a first-person view (FPV), where pilots control quadcopters at speeds in excess of 100 kilometers per hour, using a remote control together with a headset connected to an on-board device, specifically a camera.

“Physical sports are more challenging for AI because they are less predictable than board games or video games. We don’t have perfect knowledge of drone models and environment, so yes AI must learn them by interacting with the real world“, explained ua Press release scientist Davide Scaramuzza, head of the Robotics and Perception Group at the University of Zurich and one of the authors of a new study recently published in the journal Nature.

The great progress achieved by this team of researchers can be understood if we consider that only a few years ago, Autonomous drones took twice as long as human-driven drones to complete the race courseunless they relied on an external position tracking system to precisely control their trajectories.

Learn in real time

Swift’s huge advantage is that it can react in real-time to data collected by a built-in camera, like the one used by human runners. Using an integrated inertial measurement unit, it calculates acceleration and velocity, while an artificial neural network uses camera data to locate the drone in space and detect the most suitable paths along the path. This information is sent to the control unit, also based on a deep neural network, which selects the ideal action to complete the circuit as quickly as possible.

Generally speaking, Swift has shown remarkable success through this work strategy, which allows him to learn and improve in real time and in the field. In recent competitions, he achieved the fastest lap, beating the best lap of a human driver by half a second. However, Human driving showed a better adaptation to changes in the environment, which, as usual, is still a big “bottleneck”. technologically overcome AI. Autonomous drones no longer succeeded when the conditions were different from the environment in which they were trained, for example if there was too much light at the location.

Related news

According to the scientists, this important advance and new developments that enable greater flexibility in artificial intelligence will enable the creation of new, more efficient applications for environmental monitoring wave accident and natural disaster managementamong other fields.

Reference

Champion-level drone racing using deep reinforcement learning. Elia Kaufmann, Leonard Bauersfeld, Antonio Loquercio, Matthias Müller, Vladlen Koltun and Davide Scaramuzza. Nature (2023). DOI: https://doi.org/10.1038/s41586-023-06419-4