CyberRunner AI Solves a Marble Maze Faster Than a Human — with Only Six Hours' Experience
Learning as it plays, CyberRunner can complete a tilting marble maze faster than an "extremely skilled" human — even when not cheating.
Researchers at ETH Zurich have created an open source artificial intelligence model dubbed CyberRunner which, they claim, shows the first signs of superiority over humans in a game of physical skill: control of a tilting marble maze, which it learns to solve on-the-hoof.
"We believe that this is the ideal testbed for research in real-world machine learning and AI. Prior to CyberRunner, only organizations with large budgets and custom-made experimental infrastructure could perform research in this area," says Raffaello D'Andrea, professor at ETH Zurich and co-lead researcher on the project. "Now, for less than $200, anyone can engage in cutting-edge AI research."
The CyberRunner's job is a relatively simple one: to guide a marble, inside a two-axis tilting maze, to its goal. When a human performs the task, they tilt the maze one way or another using a pair of knobs; a CyberRunner uses motors for the same task. In both cases, though, it's a question of physical dexterity and cerebral thinking — and, the researchers say, CyberRunner can beat humans at the task in a first for artificial intelligences.
The model learns to play the game using visual input from a top-down camera focused on the labyrinth, using a reinforcement learning algorithm with performance-based rewards and a "critic" which picks out approaches delivering better results. This happens in real-time, the researchers note, while the model is playing the game — the more it plays, the better it gets.
After a little over six hours, the team says, CyberRunner learned to outperform the previous fastest recorded time for an "extremely skilled" human player of the same maze by over six per cent — even after the team forbade it from using self-discovered shortcuts to "cheat" and skip maze portions. This, the researchers say, is the first example of an AI outperforming a human at a game of physical skill — rather than a mental game like chess.
More information, including a pre-print of the paper detailing CyberRunner, is available on the project website; the source code and hardware design will be released under an open source license on GitHub, the researchers have promised, though at the time of writing the repository was empty.
"Once thousands of CyberRunners are out in the real-world," D'Andrea predicts of its open source release, "it will be possible to engage in large-scale experiments, where learning happens in parallel, on a global scale. The ultimate in citizen science!"