As artificial intelligence systems advance in efficiency, power economy, and greater autonomy, the conventional separated architecture of sensing and computing modules can no longer meet the demands of real-time response in complex environments. Multimodal sensing-computing devices (MSCDs) offer a promising technological pathway for embodied intelligent systems, merging multi-source signal perception, data preprocessing, and neuromorphic computation within a single physical platform. This review systematically outlines recent advances in multimodal perception and neuromorphic computing, with a focus on the contrast between unimodal and multimodal perception mechanisms, as well as strategies for multimodal data fusion and decoupling. Furthermore, it explores the structural design and cross-modal coupling mechanisms of MSCDs. Representative applications of such integrated systems are also surveyed across various domains, including embodied intelligent robots, wearable electronics, bionic prosthetics, and multimodal scene recognition. By analyzing the advantages and limitations of existing technologies, this article identifies critical directions for achieving low power consumption, high integration, and adaptive learning capabilities. MSCDs not only provide new insights into the hardware realization of artificial intelligence but also lay the solid foundation for constructing embodied intelligent systems that can perceive, interpret, and co-evolve with their environments.
Xu et al. (Fri,) studied this question.