Autonomous robot learning: what can we take for free?

In this paper we show that the environment topology can be "taken for free" and simplify the learning problems on autonomous robots. The key point is to preserve the coding of the sensory information all along the neural processing chain. First, we explain how it is possible to build a robot controller inspired from neurobiological studies that can learn to associate a particular object present in a real scene with a particular movement (sensory-motor association). Second, the previous architecture is generalized to a dead-reckoning problem. A robust representation of goal places is provided by the association of the recognition of self-learned landmarks and their angular position in a panoramic view. This system, inspired by an hippocampus model, allows our robot to retrieve a goal location after the learning of very few places in the goal vicinity. Because of the generalization capabilities of this system, the robot does not need to learn all places in its environment to be able to return from any location in an open environment to the learned location.