Superhuman AI Robot 'CyberRunner' Excels in Physical Realm, Outperforms Humans in Labyrinth Marble Game

Researchers from ETH Zurich have shown how the learning capabilities of AI are able to surpass humans not just when it comes to the computational or cerebral sense but also in the physical realm.

AI Robot 'CyberRunner' Exhibits Great Physical Skill

While there have been examples of AI beating chess, poker, or go masters in the past, researchers have taken this new achievement to the next level. In the researchers' novel innovation, the AI does not just use brain power but also exhibits physical skill, coordination, and precision.

With the famous labyrinth marble game, Thomas Bi, a PhD candidate from ETH Zurich, and Rafaello D'Andrea, a professor at ETH Zurich who also founded Kiva Systems, have developed an AI robot named CyberRunner. They carried out their research at the Institute for Dynamic Systems and Control at ETH Zurich.

Now, the researchers have shared their deep learning network and AI model that enable a system to learn about a certain game at a speed that is faster than any human.

CyberRunner: a Superhuman AI Robot

CyberRunner has two motors booted on it that allow it to play like a human. The AI robot learns through experience with recent advancements in model-based reinforcement learning. It exhibits the capacity to make informed decisions regarding behaviors that could be successful by planning for the future.

Similar to a person, CyberRunner can practice maze navigation and continually gets better in how it understands the game. As it plays the game, the robot gathers observations and gets rewarded based on performance. This is done through the camera's "eyes" that look at the labyrinth. The bot then keeps a memory of the experiences it collects.

With the memory, the model-based reinforcement learning algorithm learns about the system's behavior. Based on how it understands the game, the AI also recognizes the behaviors and strategies that show more promise.

After the bot was done practicing for six hours, it was capable of finishing the maze faster compared to any other time record. In fact, CyberRunner was able to outperform the previously fastest recorded time by 6%.

Interestingly, as the AI robot learned, it also naturally found shortcuts. The researchers had to explicitly tell the robot not to follow any of the shortcuts it discovered.

With this, CyberRunner demonstrates that the world of artificial intelligence is not just code-related but that it also involves the mastery of physical challenges.

"We believe that this is the ideal tested research in real-world machine learning and AI," Professor D'Andrea comments. "Prior to CyberRuner, only organizations with large budgets and custom-mad experimental infrastructure could perform research in this area. Now, for less than 200 dollars, anyone can engage in cutting-edge AI research."

Professor D'Andrea adds that when thousands of CyebrRunners are deployed into the real world, it will become possible to "engage in large-scale experiments, where learning happens in parallel, on a global scale." This will be the "ultimate in Citizen Science."

Check out more news and information on Tech & Innovation in Science Times.