Imagine an autonomous robot trained by a car gets into a Formula 1 car and beats Max Verstappen, Lewis Hamilton and Fernando Alonso in a series of races, setting the fastest lap times. That’s more or less what happened in the field of drone racing in controlled tests, the results of which were released this Wednesday. in the magazine ability.
In a paper led by a prestigious researcher Davide Scaramutzafrom the University of Zurich, researcher Elijah Kaufman It details how an artificial intelligence (AI) system they called Swift took on three human champions, including world champions from two international leagues, and won 15 out of 25 races. And not only that, as expected a year ago RAI chainIt also achieved the fastest race time ever recorded over the entire course, half a second ahead of the best time recorded by a human driver.
“Remember when IBM’s Deep Blue beat Garry Kasparov, or when Google’s AlphaGo beat the eventual Go champion? These competitions, in which artificial intelligence algorithms prevailed over human champions, are key milestones in the history of artificial intelligence,” Scaramuzza explained to elDiario.es. “Our result is the first time an AI robot has beaten a champion in a real physical sport designed for and by humans.”
One on one racing
The piece offers details of a test that took place in June 2022 and featured two international champions, Alex Vanover and Thomas Bitmata, and three-time Swiss champion Marvin Scheper. All three drivers received a week of practice at the race track, after which each competed against the Swift in multiple one-on-one races. In this type of competition, professionals drive high-speed vehicles through three-dimensional circuits, and each pilot sees the environment from the perspective of his drone through video transmitted from an on-board camera.
In previous trials, humans were at a clear disadvantage because the autonomous system had a device that captured motion from outside the circuit, while humans only had images on the screen. In this case, Kaufman and his colleagues wanted to make sure the playing field was as even as possible, and arranged for the Swift to use only its onboard sensors and computing power, using what’s known as “reinforcement learning.” Deep”, which consists of: a real-time simulation based on data collected in the physical world.
This is the first time an AI robot has beaten a human champion in a real physical sport designed for and by humans.
– University of Zurich
Because, as Scaramuzza points out, unlike chess or Go competitions, this is about competing in a fast and changing environment. “Swift practiced in the simulation for 50 minutes,” he says. “This corresponds to the equivalent of 23 days of real flight hours with no downtime (no sleep) in which the Swift first learns to fly from scratch and then compete. And it does this through trial and error, making mistakes and rewarding actions that minimize errors (accidents) and maximize the completion of the race in the least amount of time.”
Moving into the unknown
“The biggest limitation they put on themselves is doing everything with the processor in the car,” he explains. Julian Estevez, a professor at the University of the Basque Country (UPV/EHU) and robotics specialist who was not involved in the research. In this case, the drones are trained to distinguish between doors and non-doors and take optimal trajectories to pass through one and, if possible, focus on the other. And although the circuit is memorized, one of the main difficulties is adjusting the position and orientation of the aircraft during flight. “In the absence of cues, the further a robot moves, the less it knows where it is,” Esteves emphasizes. “At first you’re 98% sure you’re at point one, but after you’ve gone a little further and accumulated motion and sensor errors, you’re 85% sure you’re at point two, and so on ad infinitum.” .
In the absence of guidance, the further the robot moves, the less it knows where it is
– UPV/EHU professor and robotics specialist
JOSE LUIS BLANCO CLARACOA professor of mobile robotics at the University of Almería, recalls that Scaramuzza and his team in Zurich are a world reference in drone and high-speed camera vision research. Here, as the authors themselves admit, their autonomous drone has several small advantages. “it is possible they feel Its IMU (inertial system) motion on board, while humans don’t,” he lists. “And it has lower latency, meaning everything is calculated on board, whereas it takes a human to receive the video signal and the command signal to reach the drone, but it’s only a few milliseconds.” In addition, it was built for this particular track, which he believes suggests that a person may perform better the first time on a track they have never seen.
“The use of inertial sensors in this case is a significant advantage of drones,” adds Para Adrian Cario, PhD in Automation and Robotics and Professor of Artificial Intelligence and Robotics at IE University. On the other hand, in addition to the training time, the autonomous drones used in the study have a reaction time almost six times shorter than that of a human. “Another key difference is that autonomous drones are not afraid of collisions with other aircraft and completely ignore them, which gives them an advantage over a pilot who tries to avoid a collision to preserve the integrity of the aircraft,” he observed. This is similar to what was already experimentally confirmed by the autonomous aircraft of piloted combat aircraft: the human pilot wants to preserve the integrity of not only the aircraft, but also his own integrity, which forces him to control the device with less aggression.
“Feel” the movement
Regarding the technical details, Blanco emphasizes “Simultaneous localization and modeling” (SLAM). “It’s not easy to put all the necessary processing for stereo vision on board a lightweight drone,” he says. “The authors use inertial visual SLAM to estimate the motion of stereo images, combine it with ‘feel’ (IMU accelerations and angular velocities) and have a learned map of where the gate is, which they use to improve the global. is positioned within the scenario and thus will be able to better calculate the optimal trajectory. And gate detection in images is done with Convolutional Neural Networks (CNN), “the de facto standard for several years because they are the best technique. machine learning Interpreting or finding objects in pictures’.
“This result is an important milestone for robotics, artificial intelligence and beyond,” Scaramuzza defends. “And it can inspire the use of hybrid solutions based on learning in other physical systems, such as autonomous cars, drones and personal robots, in a wide range of applications.” “The development of autonomous aircraft has created a lot of scientific knowledge, not only in aircraft, but also in other types of robots and even outside the field of robotics, in applications such as augmented reality and photogrammetry,” admits Kario. “The main ones are related to supervisory or inspection tasks in an indoor environment: industrial inspection, warehouse management, surveillance in buildings or searching for people in dangerous access areas.”
In the coming years, we will see how these systems outperform human drivers in motorcycle and auto racing, and later in all other sports.
– Professor of Artificial Intelligence and Robotics at IE University
And what will be next? “In the coming years, we will see how these systems outperform human drivers in motorcycle and car racing, and later in all other sports,” says the expert. “In the case of racing, this technology is expected to be used from 2025 to increase vehicle safety and prevent accidents during motor racing. It is the same as a specialist in autonomous robot enterprises Guido de Crun An analysis article published simultaneously in the journal Nature notes that it would be interesting to test this technology in a more realistic and varied environment where the system would have to deal with external disturbances such as wind or change. light conditions. “Because they acquire sensory information faster than human pilots who rely on delayed images,” he concludes, “drones will undoubtedly eventually beat humans in these challenging conditions.”
Source: El Diario