Doodles may look crude and simple, but it’s actually pretty remarkable how much information they can convey. How is it that a couple of black squiggles on a sheet of paper are instantly recognizable as a fish? And, how is that the doodle is recognizable to virtually every human on the planet? NORAA is robotic art piece that explores that concept, and attempts to generate doodles of its own.
The NORAA robot itself, which was created by Jessica In, George Profenza, and Sam Price, is a pen plotter. It’s a completely custom design built from glass, machined aluminum, and Dynamixel smart actuators. The machine is a work of art in its own right, and the stunning design alone would be enough to secure it a spot in an art gallery. But, it’s the machine learning capability that allows it to understand doodles, and which is the real focal point of the piece.
NORAA utilizes the sketch-rnn recurrent neural network, which was built by Google’s Magenta team on TensorFlow. That was trained on the Google Quick Draw data set, which is a library of hundreds of doodles compiled for artificial intelligence training. To start, a user doodles an object, and NORAA will attempt to recognize it. It will then complete the doodle as it understands it. If it can’t understand the user’s doodle, it will find an item in its library that has similar characteristics and draw that instead. It’s an interesting look at a form of drawing that comes naturally to human, but which poses quite the challenge for an AI.