Navigate to remember: a declarative memory model for incremental semantic mapping

Biologically inspired computational techniques play a crucial role in robotic cognition. Artificial learning agents and robots that interact in complex environments must constantly acquire and refine knowledge over long periods of time. In this paper, we propose a novel recurrent neural architecture that mimics humans’ declarative memory system for continuously generating a cognitive map during robot navigation. The proposed method termed as Declarative Memory Adaptive Recurrent Model (DM-ARM), and consists of three hierarchical memory courses: (i) Working Memory, (ii) Episodic Memory and (iii) Semantic Memory layer. Each memory layer comprises a self-organizing adaptive recurrent incremental network (SOARIN) with a different learning task respectively. The Working Memory layer quickly clusters sensory information while the Episodic Memory layer learns fine-grained spatiotemporal relationships of clusters (temporal encoding). Both the memory layer learning is in an unsupervised manner. The Semantic Memory layer utilizes task-relevant cues to adjust the level of architectural flexibility and generate a semantic map that contains more compact episodic representations. The effectiveness of the proposed recurrent neural architecture is evaluated through a series of experiments. We implemented and validated our proposed work on the tasks of robot navigation.