Evolving an action-based mechanism for the interpretation of geometrical clues during robot navigation

Kamil and Jones have shown that Clark's nutcrackers (a species of crow) can exploit abstract geometrical relationships between landmarks to reach a target location. This has been interpreted as evidence for the existence of 'cognitive maps'. In the research reported here, we replicate the behaviour observed by Kamil and Jones, using methods from evolutionary robotics. A genetic algorithm is used to train a software model of the Khepera miniature mobile robot, controlled by an artificial neural network. The architecture of the network precludes the existence of cognitive maps. The robots display a range of different behaviours. We demonstrate that these depend on a process of 'active perception' and analyse the underlying computational mechanisms. On the basis of our results, we suggest that animals in the wild may indeed use active perception as a tool to exploit geometrical relationships between landmarks and that such a mechanism might complement other mechanisms, including cognitive maps.