Taking inspiration from psychologist and neurobiologist works, we have already proposed solutions to perform robot navigation in small environments (one or two rooms) based on a dynamical process linking sensory and motor information.Our robotic architecture thus codes for place in the environment (based on place cell model) and learns to link this code with a motor command. A robot learning a few couples of this kind can exhibit homing behaviour :  a competitive mechanism on these place/command associations tend to create a basin of attraction that converges to the goal place. The same mechanism can also be used for patrol behavior (trajectory learned with human interaction).


But several questions remain open for performing the same task in large environments. Autonomous robotic systems are facing measure imprecision coming from sensors and uncertainty linked with dynamic environments. Long term navigation and exploration are very complex tasks as emphased by the complexity of brain neural structures involved in their processing. Scale changes on the size of the environment in which the robot navigates imply to learn and handle much larger volume of data. Furthermore how to overcome the signal over noise ratio that tends to decrease as more information have to be learned? How to disambiguate visual information to reduce uncertainty on localisation? Robotic system aiming at an autonomous behavior have to cope with two main challenges. First, the system has to actively extract and learn robust information that it found relevant to adapt its current behaviour. Second, this kind of robotic system must be able to auto-evaluate its performances to detect failure or dead lock that might occur while interacting with a complex and dynamical environment (while local elementary decision can be correct!). Hence we can define autonomy for a robot as the ability to detect and to correct failures in its behavior.