Now he can destroy us in drone races too

Imagine a future where racing drivers, in addition to competing against humans, also compete against robotic vehicles that they control. Artificial Intelligence (AI). This scene, which seems to have fallen out of a science fiction movie, could one day become a reality. And although we are far from witnessing this, some promising developments have been made recently.

AI solutions go beyond the models that power ChatGPT, such as GPT-4, and the natural language processing systems that power Alexa. It is a discipline whose possibilities are enormous, and as we will see, it also includes the control of vehicles without human presence, for example drones capable of participating in a professional race.

Algorithms that outperform professional pilots

If we talk about artificial intelligence and sports, we also find several predecessors. This field has given us extraordinary milestones. To name a few examples, IBM’s Deep Blue computer beat Garry Kasparov in the chess game in 1997, and AlphaGo did the same with the best Go player in the world. As we can see, all achievements are related to board games, nothing more.

Faced with this reality, a group of Swiss researchers set out to test the AI off the board and concluded that a drone race would be an ideal scenario. So they set out to develop a quadcopter, a drone powered by four rotors, that would work with a series of conventional algorithms and programming methods to win the race.

After weeks of work, the team pitted an AI-powered drone against human pilots. And the results were surprising. The device managed to defeat its opponents in 15 out of 25 races. According to the researchers, this is the first time that artificial intelligence has been able to beat human champions in a real competitive sport, opening the door to new advances.

AI drone trajectory versus human pilot trajectory

This is a significant achievement that, as we say, required a combination of different approaches. On the one hand we have supervised learning, the type of training used to “teach” the drone to identify the gates to go through in competition. This is a task achieved thanks to the contribution of thousands of images used during training.

On the other hand, we have reinforcement learning, an approach that has been useful for a drone to discover the best possible route in a simulated environment and then put it into practice. in a real environment. In the image above, we can see the route chosen by the AI ​​(red) compared to that of its opponent (blue) in one of the races where the machine ended up winning.

It should be noted, at least when it comes to technology, that the techniques used in this test can only be useful in extremely controlled environments. Researchers recognize that small changes, such as increased light intensity or an opponent’s hit, can cause the drone to lose control and even crash. One obstacle has been overcome, but many still need to be solved.

Pictures: Leonard Bauersfeld

In Xataka: Google created the first watermark for AI: we already have the technology, now the most important thing is missing