Some initial experiments in self-organization and dynamic sensing

Earlier experiments with mobile robots, conducted in structured environments containing few perceptual stimuli, showed that identification of perceptual landmarks can be achieved by using Kohonen networks. Subsequent experiments on route learning using Kohonen networks confirmed what the initial experiments had suggested: for complex navigation tasks, over-generalization (leading to perceptual aliasing) and under-generalization (leading to failure in identification) are issues that need to be tackled. This paper presents initial experimental results on how to address these problems by means of dynamic sensing. There are two main aspects, concerning the task of landmark identification: recognition and rejection. Firstly, identical locations need to be recognized reliably as identical, despite perceptual variations in each visit. Secondly, similar looking, but distinct locations need to be differentiated, i.e. false “recognitions” need to be rejected. To achieve the former, generalisation is required. To achieve the latter, specificity is required-there is a conflict between over-generalisation and under-generalisation which, in the spatial domain, can only be resolved by fine-tuning of parameters, depending on the particular application