It's remarkable to watch the CyberRunner robot rapidly advance from shaky beginner to smooth, seasoned pro – and, like humans, it even finds ways to cheat. A compelling demonstration of AI's ability to master complex real-world skills.
Just a couple of months after we saw AI-piloted drones absolutely dominate human world champions in a circuit race, ETH Zurich researchers have revealed another real-world physical demonstration of AI's phenomenal learning capabilities.
The target here: a rolling marble maze game. You've played these things; a pair of knobs allow you to tilt the board as you roll a marble through a wooden labyrinth, attempting to avoid falling in holes.
The CyberRunner AI robot was furnished with two electric motors to drive the knobs, a camera to see the board, and an AI brain running a "deep reinforcement learning system," capable of sending up to 55 control instructions per second. It was told the goal of the exercise, and left to experiment on its own.
Naturally, early attempts were jerky and wobbly, but 6.06 hours' worth of training later, the AI smashed the fastest recorded time by an "extremely skilled human player," by more than 6%, demonstrating a smooth, practiced confidence that's mesmerizing to watch:
It's fun to note that the AI quickly identified shortcuts it could use to cut out whole sections of the maze and go even faster – the researchers had to get in and specifically tell it not to cheat.
The CyberRunner team says it'll soon open-source the entire project, making it something anyone can build and train at home.
“We believe that this is the ideal testbed for research in real-world machine learning and AI," says Professor Raffaello D'Andrea in a press release. "Prior to CyberRunner, only organizations with large budgets and custom-made experimental infrastructure could perform research in this area. Now, for less than $200, anyone can engage in cutting-edge AI research, as applied to the physical world. Furthermore, once thousands of CyberRunners are out in the real-world, it will be possible to engage in large-scale experiments, where learning happens in parallel, on a global scale. The ultimate in Citizen Science!”
This is just a cute little game, but it's a demonstration of something much bigger. 2023 has been an absolute watershed moment and a tipping point for AI systems, and their convergence with the world of robotics has driven equally shocking progress in the embodied world. It feels like a macabre privilege watching these machines learn to learn in real time, becoming ever more capable as well as faster to pick up new skills, whether from scratch or by watching others.
As 2023 comes to a close, a far-reaching technological revolution seems like a foregone conclusion at this point. The seeds of human obsolescence have been sown, and it's now a matter of just how quickly the value of both human intelligence and human labor drop close to zero. It doesn't feel like it's 20 years away; even 10 years of development at this wildly accelerating pace seems like too long.
What an extraordinary time to be alive.
Source: CyberRunner
