Comparison of Optimization Algorithms for the Indirect Encoding of a Neural Controller for a Soft Robotic Arm

With their dexterity, robustness and safe interaction with humans, soft robots bode to revolution the field of robotics. However, featuring structures undergoing nonlinear deformations and under-actuated mechanisms, traditional control techniques are usually unsuccessful. Artificial neural networks have instead shown to be a suitable solution to control soft robots in several cases. Among the different classes of algorithms to train neuro-controllers, one that recently experienced a wide spread consists of optimization with genetic algorithms (GA) through indirect encoding. Main advantages are: the ability to produce networks with functional regularities that exploit the geometry of the domain; the decoupling of problem complexity from its resolution. The predominant use of GA has several reasons, ranging from bio-inspiration to some undeniable technical advantages. However, two main issues suggest the need to explore different and possibly more efficient algorithms to train neuro-controllers for soft robots: the high computationaI cost of mathematical models to simulate soft robots and evidences of unsuccessful global convergence of GA if not carefully tuned. In this study, we compared the performance of GA with those of other optimization algorithms in training an artificial neural network to control a soft robotic arm inspired by the octopus, simulated through a non-linear dynamic mathematical model.