After strategy games such as chess and Go, robots are now also learning to play skill games. To respond to the game “Jenga”, where a tower of wooden blocks needs to be modified, provided scientists a robot arm with the sensitive touch sensors.

such A robot could perform tasks in which skill is in demand, writes the Team led by Nima Fazeli from the Massachusetts Institute of Technology (mit) in Cambridge, in the journal “Science of Robotics”.

“to conclude through buttons in the animal Kingdom is ubiquitous, but in the robot handling under-developed,” write the researchers. This information conveyed by the tactile sense could be, in the industry. “In a production line for mobile phones the impression of a snap fastener or a threaded screw comes in almost each individual step rather of power and touch than by sight,” explains co-author Alberto Rodriguez.

The researchers combined camera shots with the touch sensors and dined with these data, an artificial neural network. To shorten the learning time, was specified to the robot that he summarizes operations with the same or a similar result – for example, the Collapse of the Jenga-tower – in clusters together. In this way, the robot had a steeper learning curve than with other methods, and it only took about 300 Attempts, instead of tens of thousands to achieve a good game result.

Initially, the learning behavior of the robot was simulated. Then the Play was followed by cubes with real Jenga. When the robot arm began the Play, he tested randomly selected wooden building blocks. He moved about a Millimeter and evaluated them according to categories such as “easy to move”, “move” and “not to move”. Depending on the assessment, he continued his efforts to remove the block – or not.

Removed components are placed at Jenga at the top of the tower. The robot managed to after a short learning time, to remove 21 or more blocks and to place new, without the tower tipped. “We have seen how many blocks could pull a man out before the tower fell, and the difference was not so great,” says Miquel Oller, another author of the study. The aim of the researchers is not it but in the end, to make the robot an unbeatable Jenga master. You want to explore with the combination of visual and ertasteten data gained skills.

In another article in “Science-Robotics” deals, Robin Murphy of Texas A&M University with the representation of robot learning in the Science Fiction books and movies. Common motifs are in the process of that Learning for robot is easy and that it leads to sentience. Both were in reality not so, writes Murphy. A thing going but it is mostly correct: That it is very difficult to teach a robot to learn the Right thing to do.