Approximating the visuomotor function for visual servoing

This paper introduces a new approach to visual servoing by learning to perform tasks such as centering. The system uses function approximation from reinforcement learning to learn the visuomotor function of a task which relates actions to perceptual variations. The function model is linear and tile coding is used for generalization. The gradient-descent SARSA algorithm is used to determine the parameters. Experiments show that the system learns to center targets at different depths with stereo vision and fully reconfigures itself in monocular case.