Researchers at the Massachusetts Institute of Technology’s Mechanical Engineering laboratory have created a robot arm that can see a Jenga tower, sense the state of its blocks, and choose which one should be taken out next.
Developed by Alberto Rodríguez, assistant professor of mechanical engineering and member of MIT’s MCube Lab, the robot doesn’t use a traditional AI learning method. Instead of depending on adversarial networks working on gigantic datasets to learn and then take decisions in the real world, his machine learns on the fly using a hierarchical model.
“This model allows the robot to estimate the state of the piece, simulate possible moves and decides on a favorable one,” the MIT researcher explains, “allowing for a gentle and accurate extraction of pieces.” His machine learning method uses visual sensors to “divide the tower in clusters, each with its own sets of physics”.
Then, pressure-sensitive touch sensors precisely measures if a piece is less or more stuck, passing on dubious moves until it finds the optimal one.
The most fascinating fact, however, is that the robot teaches itself how to play Jenga in the best way possible, getting better and better the more it plays.
According to Rodríguez, this tech is not just for gaming: “The technology could be used in robots for manufacturing assembly lines.” It makes sense but what I really want to know is if I would be able to play Twister with Atlas — before it kills me.