Researchers at ETH Zurich have created an open-source AI model called CyberRunner that shows signs of superiority over humans in a game of physical skill.
CyberRunner's task is relatively simple: guide a marble in a maze with two axes of inclination toward a target. When a person performs this task, they tilt the maze to one side or the other using their hands. CyberRunner uses motors to accomplish the same task. Both involve physical dexterity and cerebral thinking. According to the researchers, the CyberRunner can outperform humans at this task, marking the first AI achievement.
The model learns to play the game using a visual cue from a top-down camera pointing at a maze using a reinforcement learning algorithm based on performance rewards and "critiques" that leads to the selection of the best score. The researchers note that everything happens in real time. The more the robot plays, the better the result.
According to the team, after just over six hours, CyberRunner outperformed the result of an experienced player (human) by more than 6%. According to the researchers, this is the first example of AI outperforming humans in a game that requires physical skills rather than mental skills.
The model learns to play the game using a visual cue from a top-down camera pointing at a maze using a reinforcement learning algorithm based on performance rewards and "critiques" that leads to the selection of the best score. The researchers note that everything happens in real time. The more the robot plays, the better the result.
According to the team, after just over six hours, CyberRunner outperformed the result of an experienced player (human) by more than 6%. According to the researchers, this is the first example of AI outperforming humans in a game that requires physical skills rather than mental skills.