Neuromorphic Mechatronic System for Real-Time Robotic Reflex Generation Using Spiking Neural Networks
Abstract
Neuromorphic computing offers a promising avenue to replicate human-like intelligence in energy-constrained, real-time applications such as edge AI and cognitive robotics. However, existing models like classical Spiking Neural Networks (SNNs) and Evolving Connectivity designs exhibit limitations, including high latency, moderate accuracy, and poor energy efficiency, making them less suitable for time-sensitive decision-making. This study introduces a novel, biologically inspired learning framework that integrates Memristor-based neural computation with reinforcement learning to overcome these challenges. The primary aim is to develop an energy-efficient, real-time capable, and scalable SNN architecture that mimics brain-like processing. The core objective lies in enhancing synaptic plasticity using reward-modulated learning combined with memristive device simulation, thus improving precision and adaptability. The novelty of this work lies in the tight coupling of neuromorphic hardware (memristors) with reinforcement strategies to achieve faster learning with reduced computational overhead. The entire system is implemented using Python and the Brian2 simulator and evaluated on a custom spiking dataset in a robotic control scenario. Experimental findings demonstrate that the proposed method achieves 93.20% accuracy, very high energy efficiency, 7.4 ms latency, and full compatibility with memristive hardware. Comparative evaluations show superior performance over NBox Memristor SNN and Spikingelly Framework, establishing its promise for next-generation neuromorphic control systems.