Neural network robotic control of unknown mass payloads

Summary form only given. The authors discuss a neural-network-based controller suitable for position control of a one-degree-of-freedom robotic manipulator with unknown mass payload. A feedforward neural network (NN) is utilized to learn the dynamics of the manipulator and provide a drive signal, based on NN inputs, to a proportional-derivative controller used to stabilize the plant. The NN estimates the payload mass implicitly using readily available state information during training and operation. Computer simulations are used to assess the NN controller performance and to compare the performance to that of the linear controller. Specifically, the NN controller is trained on one trajectory for three different payload masses using the measured actuator torque at a given state as an estimation of the payload mass for input to the neural network