The purpose of this work is to provide a robust vision-based input device. In our system, a programmable retina is looking at the user who sends commands by moving his hand. The fusion between the acquisition and the processing functions of the retina allows a close adaptation to the lighting conditions and to the dynamic range of the scene. Thanks to its optical input and massive parallelism, the retina computes efficiently the contours of the moving objects. This feature has nice properties in terms of motion detection capabilities and allows a dramatic reduction in the volume of data to be output of the retina. An external low-power processor then performs global computations on the output data, such as extreme points or geometric moments, which are temporally filtered to generate a command.