Connecting Experiences: Towards a Biologically Inspired Episodic Memory for Developmental Robots

Abstract

Robots must flexibly connect context-relevant chunks of past experience to meet novel goals. We introduce a neural framework that encodes high-dimensional sensorimotor episodes into structured episodic memory and retrieves relevant sequences to simulate unexperienced actions.

Methodology

The model uses self-organizing maps to discretize sensory and motor inputs, linked via a hub layer. Episodes are stored as trajectories in this low-dimensional space. A cue triggers hippocampal-like recall through spreading activation and a dynamics simulates future states. Robots must flexibly connect context-relevant chunks of past experience to meet novel goals. We introduce a neural framework that encodes high-dimensional sensorimotor episodes into structured episodic memory and retrieves relevant sequences to simulate unexperienced actions.

The model uses self-organizing maps to discretize sensory and motor inputs, linked via a hub layer. Episodes are stored as trajectories in this low-dimensional space. A cue triggers hippocampal-like recall through spreading activation and a dynamics simulates future states. Robots must flexibly connect context-relevant chunks of past experience to meet novel goals. We introduce a neural framework that encodes high-dimensional sensorimotor episodes into structured episodic memory and retrieves relevant sequences to simulate unexperienced actions.