One of the main claims of active vision is that finding data on demand, based on the requirements of the task, is more efficient than reconstructing the whole scene by performing a complete visual scan of it. This aids generalisation and a dramatic reduction of the needed visual computations. Using this strategy, however, generates the need to learn complex gaze control strategies dependent on the pursued goals and the properties of scenes and objects. For example, to be able to find an object in the environment an agent needs to learn to use several sources of information such as spatial relations of objects and bottom-up saliency of scene regions. In addition, if the system is genuinely autonomous it also needs to develop a representation of the objects themselves, for example of potential targets, cues and distractors, on the basis of generic reward signals to be maximized and the visual control policy used. Most of the models proposed in developmental robotics do not use adaptive visual control and so are ill suited to investigate these issues. In a previous work we presented a reinforcement-learning neuro-robotic architecture, based on neural population codes, which was able to develop attention control policies by interacting with the environment based on a rewarded reaching task it had to accomplish. In this paper the same architecture is used to investigate the types of internal representations that this same architecture develops when exposed to two classes of environments where objects are organised on the basis of contrasting spatial relations.
