Abstract: To achieve human-like autonomy and adaptability in complex unstructured environments, robots must undergo a paradigm shift in multimodal perception systems by drawing inspiration from neuroscience. However, existing studies often remain at superficial descriptions of biological mechanisms, failing to deeply demonstrate how these principles can systematically guide innovation in robotic perception theory and technology. This review aims to bridge this gap by centering on insights from neuroscience to systematically construct a technological blueprint ranging from bio-inspired sensors to brain-like fusion algorithms. First, this review provides an in-depth analysis of the neural circuitry underlying multisensory integration in the brain, extracting engineering-ready computational principles. It then systematically examines cuttingedge sensor technologies such as neuromorphic vision and flexible electronic skin, emphasizing their applications in tasks like real-time state estimation. At the algorithmic level, the review focuses on how deep learning techniques, including variational autoencoders, cross-modal attention mechanisms, and spiking neural networks, can be employed to implement predictive coding and active perception in robotic systems. Finally, the article critically discusses core challenges in translating neuroscientific inspiration into engineering practice and outlines future directions such as neuromorphic computing, standardized embodied datasets, and machine self-body awareness. This work provides a clear development pathway for building next-generation embodied intelligence systems capable of genuine environmental perception and interaction.

Key words: Multimodal perception')">Multimodal perception, Robotics, Neuroscience-inspired, Embodied intelligence, Neuromorphic computing, Predictive coding