Evolving Neural Networks for Robotic Arm Control

Abstract

Developing effective and adaptive robotic arm controllers is crucial for many industries, e.g. manufacturing. Traditional pre-programmed controllers cannot adapt to changing environments. This study investigates how neuroevolution can be used to develop robotic arm controllers and addresses key gaps in the existing literature, such as incorporating expert demonstrations and analyzing the robustness of evolved controllers. In addition to addressing these questions, this work compares different controller architectures and training algorithms. The proposed evolutionary neural network motion controller can accurately complete the random target reaching task, moving to within 1.7 cm from the target on average. An evolutionary supervisor neural network approach is also proposed to solve the pick-and-place task. The proposed method achieves a high successful completion rate, 927 out of 1000 trials.