Two-Legged Robot System Identification With Artificial Neural Networks

Legged robot platforms, unlike wheeled robots, have a hybrid dynamic structure consisting of the flight and contact phases of the legs. Since the hybrid dynamic structure and nonlinear dynamics in the robot model make it difficult to calculate the system Jacobian, the implementation of adaptive control methods may become harder. In this study, the effects of different neural network architectures and neuron models in the system identification of hybrid dynamical systems are investigated. Two-legged robot system is used as the hybrid dynamic system. Due to supervised learning technique to be used in the training of neural networks, the central pattern generator has been utilized as a controller to generate walking datasets. Afterwards, neural networks with 2, 4 and 5 layered feedforward and recurrent neural models were trained with datasets using serial-parallel system identification architecture.