You know the problem: all of those LEGO kits you’ve bought over the years eventually end up all mixed up in a big tub. Before long, you’ve got a bin full of thousands of different kinds of bricks, and finding any specific kind of brick is like searching for a needle in a rainbow stack of needles. Fortunately, Paco Garcia has developed a solution, and you can follow his lead to build your own automatic LEGO sorter with TensorFlow running on a Raspberry Pi.
Garcia was inspired to build his LEGO sorter after seeing how a Japanese farmer used TensorFlow to sort cucumbers. The basic concept is straightforward: a camera takes a photo of a LEGO brick, and then TensorFlow determines what kind of brick it is. The brick is then placed in a bin with others of its kind. Of course, actually executing that concept was a big challenge, and Garcia’s automatic sorting machine took a lot of mechanical engineering skill.
The LEGO bricks start on a chute as a group, and then slide down onto a slow moving conveyor to thin out the group. Next, a second, faster moving, conveyor further separates the bricks. With a decent space between one brick and the next, they’re funneled into an area where they can be photographed. The Raspberry Pi with TensorFlow then decides what kind of brick it is and rotates another chute to the corresponding bin, where the brick is deposited. Garcia’s first run of the machine accurately sorted 89% of the bricks, and that can presumably be improved with better TensorFlow training.