PaintRL: Coverage Path Planning for Industrial Spray Painting with Reinforcement Learning

We present PaintRL, a framework that enables research on optimizing industrial spray painting for arbitrary objects with reinforcement learning. PaintRL implements a toolkit to simulate and visualize spray painting based on the physics engine PyBullet. By means of this toolkit, we train neural networks to predict coverage paths and evaluate the results on two objects: a quadratic sheet and a real car door. Our initial results show that the generated coverage path of the car door performs on a par with a manually implemented zigzag baseline. To allow sim2real transfer for non-resettable tasks like spray painting, we replace paint by light using projection mapping. This approach opens up new possibilities to visualize the results from simulation, collect human demonstrations and capture real-world images. PaintRL is part of our endeavor to utilize the recent advances in deep reinforcement learning for economically important industrial tasks.