Edge AI for Low-Power IoT Devices: Architectures, Algorithms, and Applications

Abstract

The convergence of Artificial Intelligence (AI) with the Internet of Things (IoT) has given rise to Edge AI—a paradigm that enables real-time, intelligent processing on resource-constrained devices deployed at the network edge. Unlike traditional cloud-based systems, Edge AI eliminates the need for constant connectivity, reducing latency, preserving privacy, and enabling mission-critical responsiveness. However, deploying AI models on low-power IoT devices, such as microcontrollers and sensor nodes, introduces significant challenges due to limited computational resources, energy constraints, and memory overhead.

This paper presents a comprehensive literature review on the state-of-the-art developments in Edge AI for low-power IoT devices up to 2021. We analyze lightweight neural architectures (e.g., TinyML, MobileNet, SqueezeNet), hardware-aware model optimization techniques (quantization, pruning, and knowledge distillation), and dedicated edge hardware platforms (e.g., ARM Cortex-M, Google Edge TPU, NVIDIA Jetson Nano). The paper also discusses software frameworks like TensorFlow Lite Micro and ONNX Runtime that support efficient model deployment on ultra-low-power devices.

Further, we review notable applications across domains such as smart healthcare, predictive maintenance, smart agriculture, and autonomous sensing. The survey highlights ongoing challenges, including real-time inference under strict energy budgets, security at the edge, and lack of standardized benchmarks. We conclude with open research directions that emphasize the need for co-optimized hardware-software design, federated learning, and scalable edge intelligence for next-generation IoT ecosystems.