Design and Comparison of Two Evolutionary and Hybrid Neural Network Algorithms in Obtaining Dynamic Balance for Two-Legged Robots

Abstract

In a closed-loop control scheme, the proportional–integral–derivative (PID) controllers are mainly used to diminish the error between the desired setpoint and the actual measured value. It is significant to note that some kind of tuning procedure is essential to acquire the optimal values for the gains of the controller to reach the desired setpoint. In this chapter, to tune the gains of the PID controller, the authors are made to develop an adaptive neural network algorithm. Further, the authors of this manuscript proposed a metaheuristic optimization algorithm, that is, modified chaotic invasive weed optimization (MCIWO) algorithm to train the weights of the neural network (NN). Moreover, the well-established optimization algorithm, that is, particle swarm optimization (PSO) algorithm has also been separately used to optimize the structure of NN. Once the controllers are developed, the required torque to move every joint of the two-legged robot on the said terrain will be predicted. In addition to the above parameters, the zero moment point (ZMP) of the foot and dynamic balance margin (DBM) of the gait have also been calculated and considered as a measure to compare the performances of the developed approaches. Further, the neural network-based PID controller tuned with MCIWO algorithm is seen to produce more dynamically balanced gaits than that of the PSO trained NN. Finally, the optimal gaits obtained through the MCIWO-NN algorithm have been confirmed on an actual two-legged robot in our laboratory.