Improving Positioning Accuracy of an Articulated Robot Using Deep Q-Learning Algorithms

Positioning accuracy of articulated robots decreases significantly when they are fully loaded and operated at maximum speeds due to increased inertia. Hard-coding correction algorithms using traditional methods is extremely difficult. A system, which could automatically detect patterns in deviations and offer possible corrections at every motion cycle would be preferable. This work explores the possibility to use deep q-learning algorithms to achieve this. Around forty experiments of various lengths were conducted. They were divided into three experimental groups, each of which had various parameters values and elements of algorithms. While algorithms in two experimental groups were unsuccessful in achieving improved accuracy, one offered comparable accuracy, while resulting in more stable and predictable deviations compared to uncorrected positioning.