Auto gain-tuning for trajectory following control based on neural network

Serial link manipulators have been used for various situations in industry, i.e. welding, assembling, painting, polishing and machining. In machining by the manipulator, material is must be soft such as polystyrene foam because the rigidity of the manipulator is small. Recently, polystyrene form machining by the manipulator is desired to make a lost pattern used in full-mold casting. In the machining operation, a trajectory following control is usually used because the desired trajectory can be generated from a cutter location data. When a manipulator is controlled to cut the polystyrene along the designed shape, various desired trajectories are generated in terms of robot coordinate system. To reduce the positioning error of the manipulator, suitable feedback gains of the controller should be calculated for each trajectory. However, the calculation of the feedback gains consumes much time because it needs to solve a non-convex optimization problem. Therefore, faster calculation method of suitable feedback gains for each trajectory is desired. In this study, we propose an auto gain-tuning method of feedback gains using neural network with training. A training set has six training pairs prepared in advance, which consists of displacement of manipulator as input and optimized feedback gains as target. In the preparation, the feedback gains for the each trajectory calculated by the 4-1-4 polynomial interpolation is optimized by using the genetic algorithms. The neural network are trained by updating the synaptic weights with back propagation. After the training, the neural network can output suitable feedback gains for any trajectories. The effectiveness of the proposed method is confirmed through 1,000 times dynamic simulations with randomly generated trajectories.