Since the publication of the “Review of Embedded Artificial Intelligence Research” in 2023, driven by innovations in hardware architectures, advances in lightweight algorithms, and the maturation of edge–cloud collaboration technologies, embedded artificial intelligence (embedded AI) has progressed from “technically feasible” to “large-scale deployment”. As a continuation of that review, this article systematically surveys the core advances in embedded AI from 2023 to 2026. At the hardware level, it examines engineering progress in non-von Neumann architectures such as compute-in-memory and neuromorphic chips, as well as heterogeneous integration technologies. At the algorithmic level, it covers dynamic adaptive lightweighting, specialized edge-side optimization of large models (including on-device large language model fine-tuning and edge diffusion models), and lightweight multimodal approaches. In terms of deployment paradigms, it discusses edge-side full training, federated edge learning, edge–cloud collaborative intelligence, and emerging paradigms. At the application level, it illustrates the “perception–decision–execution” pipeline in industrial IoT, wearable healthcare, autonomous driving, embodied intelligence, and smart agriculture. The article also analyzes core challenges including ultra-low-power design for extreme scenarios, cross-platform standardization, edge-side data security and privacy, and model robustness in complex environments. Based on these findings, four research directions are proposed to guide future work.
Zhaoyun Zhang (Sat,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: