Speeding-Up the Learning of Saccade Control

A saccade is a ballistic eye movement that allows the visual system to bring the target in the center of the visual field. For artificial vision systems, as in humanoid robotics, performing such a movement requires to know the intrinsic parameters of the camera. Parameters can be encoded in a bio-inspired fashion by a non-parametric model, that is trained during the movement of the camera. In this work, we propose a novel algorithm to speed-up the learning of saccade control in a goal-directed manner. During training, the algorithm computes the covariance matrix of the transformation and uses it to choose the most informative visual feature to gaze next. Results on a simulated model and on a real setup show that the proposed technique allows for a very efficient learning of goal-oriented saccade control.