A Bio-realistic Synthetic Hippocampus for Robotic Cognition

Abstract Current robotic systems struggle with adaptive generalisation beyond curated training domains. Inspired by hippocampal dynamics in biological cognition, we introduce a synthetic memory architecture that segregates online sensorimotor interaction from offline consolidation and generative replay. Implemented via spiking neural networks and neuromorphic substrates, our framework enables bidirectional memory traversal, goal-prioritised plasticity updates, and energy-efficient policy synthesis. This dual-state system bridges real-time control with autonomous learning, advancing a biologically grounded pathway toward resilient, context-adaptive robotic intelligence.