Artistic Painting Robots

There are few jobs that men are better at than robots. After all, robots are more precise, they have superb physical conditions, they do not complain on repetitive jobs, they are never on strike. But there is one sector that is undoubtedly a no-robots zone: the creative industries. Robots do not understand the spontaneity of creativity, the artist’s sensitive eye or even the basic principles like ugly or beautiful...

But what if a robot can really learn? What if it is capable of mastering painting techniques, and by taking the public opinion on beautiful and ugly, can create trend-adapting, ever original paintings we love? This would revolutionize the art world completely.

Imagine on your next museum visit, that just by giving more or less attention to a painting, you could actively change the type of art that is created? Such a robot will not only make exhibitions extremely user centered, but could also function as a research tool, capturing visually and in-time the changing taste of its visitors.

Robot’s art school

Within the course of this two week module, we changed an ordinary 3-axis robot (a CNC milling machine) into a learning artist. We developed our own algorithm, based on Q-learning to teach the robot about our taste.

We let the robot paint 75 different paintings. One painting would take 4 actions, that the robot could choose himself from following options:

  • Taking yellow paint
  • Taking blue paint
  • Drawing a rectangle (random size and position)
  • Drawing a circle (random size and position)

  • A lot more actions were implemented in the coding platform (like drawing lines, different pressures on the brush, etc), but for this first trial the robot was given a limited amount of options to make the learning progress clear.

    The Artistic robot at work


    To make feedback less ambiguous, some rules of ‘our likes’ were fixed: we would train the robot to make blue rectangles, yellow squares and preferably to put variation on a canvas (so different colours and different shapes).

    By looking at all 75 paintings that the robot made the learning can clearly be seen. In the beginning the order of the actions are random, multiple canvases are empty since the robot did not know that it needed to pick up paint before drawing. Towards the end of the process, you see more and more paintings that have both yellow and blue, and that have blue squares, and yellow rectangles, proving the system to actually learn how to paint with the rules we set before.

    All 75 paintings


    The robot was the outcome of the two-week module ‘Learning Robots’ by Emilia Barakova, which aims at providing the necessary tools and knowledge for making products able to learn and behave autonomously, being able to show elements of natural behavior. This robot was developed in a team together with Thomas van Lankveld and Frits Stam.