Neuromorphic Computing and Spiking Neural Network Hardware Design for Autonomous Control Systems

Abstract

Autonomous control systems, including robots, unmanned aerial vehicles (UAVs), and smart vehicles, require rapid, efficient, and adaptive processing to perceive the environment and act upon it in real time. Neuromorphic computing, which mimics the brain’s neuron architecture using spiking neural networks (SNNs) and event-driven hardware, presents one significant alternative to von Neumann and deep learning approaches for these applications. This paper presents a detailed technical review of design strategies for neuromorphic hardware incorporating SNNs, as well as its use within autonomous control systems. We consider the architecture of SNNs on silicon—including digital vs. analog designs, on-chip communication, synapse memory, and on-chip learning. The benefits of neuromorphic SNN hardware—such as massively parallel, event-driven computation, and low-power consumption—are compared to conventional v. von Neumann architectures and deep learning. The implications of these benefits are summarized for real-world autonomous control systems in terms of ultra-low-latency sensorimotor loops and energy-efficient continuous operation. Figure 1 depicts an example control architecture that uses a generic neuromorphic processing unit to conduct an SNN. Figure 2 presents a case study of an SNN system for navigation. Table 1 summarizes our comparisons of architecture between traditional and neuromorphic. We review experimental demonstrations and simulation studies using neuromorphic hardware platforms conducting autonomy tasks especially on applications such as obstacle avoidance, adaptive flight control, and end-to-end autonomous driving. The results from these studies demonstrate SNN-based controllers match or exceed classical controllers while providing a significant decrease in power requirements. Finally, we explore topics and challenges facing neuromorphic computing—programming complexity, training algorithms, and integration. We discuss emerging topics that are likely growing the future of neuromorphic hardware within next-generation autonomous systems in robot and vehicle applications—advancements in SNN training algorithms, developing new memristive synaptic devices, and larger systems and processors. The potential implications regarding hardware for autonomous systems is substantial, meaning neuromorphic SNNs will be increasingly used in future autonomous robots and vehicles more resiliently, whilst being more implanted to all electrical machine learning approaches that have a much closer inspiration to bio-systems.