Researchers at MIT have created a robot that is learning how to play the popular game Jenga. The game involves stacking rectangular bricks in layers perpendicular to each other, and one-by-one removing a block to place on top of the tower. Once the tower falls, the game is over.
This robot has learned the game by grouping outcomes of its moves into clusters. This has allowed the researchers to have a successful machine while avoiding thousands of attempts, which would mean building the tower thousands of times.
Machine learning has most commonly been used for intellectual tasks so far. The general public is widely familiar with these actions, such as Amazon recommendations based on past purchases, G-Mail suggestions when writing emails, or detecting fraud in online purchases. This development is relevant because it is the first step into machine learning performing physical tasks. The Jenga-robot uses visual and physical cues in addition to knowledge and data. Data is gathered based on feedback from the machines camera and claw to correlate the position, tension, and movement of the piece with the result of the move.
With the success of this Jenga-playing robot, machine learning can venture into more physical tasks, like separating items in a recycling center. The research will also help with the recognition of physical clues that humans instinctively know. The snap we feel when closing a container is not easily persevered visually, but it is the main indicator that the container has been successful sealed. These advancements will help to expand the scope of machine learning and broaden its reach into even more industries.