THE race is on. Two miniature cars are zooming around a loop of track in a new robotics arena. I鈥檓 driving one via remote control. The other is controlled by artificial intelligence 鈥� and it鈥檚 kicking my ass.
From August 2017, the University of Essex鈥檚 robotic games laboratory will pit AI systems against each other in robot challenges, starting with these remote-controlled cars. The AI I鈥檓 racing has been honed to complete the track in super-quick time, but the real competitions won鈥檛 be so straightforward. The rules will change between heats: in one, the cars鈥� brakes might be disabled, in another, obstacles could be added to the track. The AI developers won鈥檛 know in advance what challenges they will face. Instead, their systems will have to figure out for themselves how to win each time.
鈥淭he challenges will be wide and varied, so the entries will have to be able to adapt to whatever tasks we present to them,鈥� says , the head of the lab.
Advertisement
For the past three years, Lucas鈥檚 department has hosted the , which is sponsored by Google鈥檚 DeepMind. This pits research teams against each other to develop intelligent systems that can take on a variety of different video games, from puzzles to platformers like Super Mario Bros. This new lab takes the concept out of simulated environments and into the real world.
Now the AIs aren鈥檛 just playing with pixels; they are controlling physical robots. The arena has 26 infrared cameras that track movements at a rate of 250 times a second, allowing the competing systems to know the precise location of the robot they are controlling at any given time.
鈥淣ow they are not just playing with pixels; they are controlling physical robots鈥�
The systems that do well will need skills that are in high demand in everything from video games to autonomous cars.
The idea is to use robot games to develop AI systems with more general capabilities. Plenty of AIs excel at very specific tasks, like beating a human at chess or recognising a car in an image, but they fall to pieces if asked to do something outside of their training. An ability to adapt is especially important for systems with applications in real-world environments, where situations are often unpredictable.
鈥淕ames are great at challenging our brains and forcing us to learn certain skills,鈥� says at New York University. 鈥淭he same goes for AI. With the right games it can slowly figure out how to do more complex tasks.鈥�
Toy cars are just the beginning. The arena can also handle other types of robot. In the future, Lucas plans to have an aerial event in which AIs will control drones.
The stakes are high. 鈥淭he AIs in the robotics competition will have to learn very quickly,鈥� says Lucas. 鈥淚f you make a fatal mistake in a video game you can just hit reset. That鈥檚 not true in the real world.鈥�
This article appeared in print under the headline 鈥淩obot game arena puts AIs to the test鈥�
