Monolithic Integration of Neuromorphic Electronic Skin and In-Memory Computing Chips for Fine Manipulation in Embodied Intelligent Agents

Abstract

For fine manipulation in embodied intelligent agents, tactile systems are required not only to provide highly sensitive perception, but also to support low-latency information processing and highly integrated hardware implementation. To address the large data-movement overhead, delayed response, and limited compactness caused by the separation of sensing, transmission, and computation in conventional tactile architectures, this work proposes a monolithically integrated platform that combines neuromorphic electronic skin with an in-memory computing chip to establish an efficient task-oriented tactile information processing pathway. Within a unified hardware architecture, the platform couples mechanical stimulus perception, neuromorphic encoding, and on-chip computing, enabling external tactile inputs to be dynamically represented at the sensory source and further processed for feature extraction and pattern discrimination in the in-memory computing array. The structural design, operating principle, and electrical, neuromorphic, and computing characteristics of the integrated platform are systematically investigated. In addition, representative fine-manipulation-related tasks, including contact-state recognition, slip warning, and texture classification, are employed to validate the system. The results demonstrate that the proposed platform effectively preserves the temporal information embedded in tactile events and outperforms conventional discrete architectures in terms of task-processing accuracy, real-time capability, and data-movement efficiency. This study provides a promising route toward highly integrated, low-latency, and task-driven tactile intelligence hardware, and lays the groundwork for the co-design of tactile sensing and on-chip processing in next-generation embodied intelligent systems.