Abstract This paper presents an investigation on the trajectory control of a robot using a new type of recurrent neural network. A three-layered recurrent neural network is employed to estimate the forward dynamics model of the robot. Standard backpropagation (BP) algorithm is used as a learning algorithm for this network to minimise the difference between the robot actual response and that predicted by the neural network. This algorithm is employed to update the connection weights of the neural network controller with three layers using a gradient function.