Unsupervised Neural Architecture for Sensorimotor Mapping in Perceptually Aliased Environments

This paper addresses the challenge of autonomous navigation in environments with perceptual aliasing, where observations are not unique; posing difficulties for current Simultaneous Localisation and Mapping (SLAM) systems. The importance of developing cognitive maps inspired by the hippocampal/entorhinal system (H/E-S) for spatial and relational memory tasks for intelligent behaviour and flexible navigation is discussed. The paper introduces the Merge Expand when Required Clone Structured Representation Yielding explainability (MERCURY) network, an unsupervised neural architecture that learns sensorimotor maps in aliased environments through continuous self-organisation. Experimental results demonstrate MERCURY's improved performance in mapping aliased environments compared to other approaches. The paper concludes with a discussion on the limitations and directions for future research to enhance the robustness and applicability of the proposed approach.