A Framework for Embodied Intelligence Assembly Robots Co-driven by Large-scale Model and Small-scale Models

Modern aerospace manufacturing, exemplified by satellite assembly, encounters significant challenges in precision, adaptability, and safety stemming from complex, low-volume production demands. Conventional robotic systems are hindered by dynamic constraints, inadequate flexibility, and inefficient safety responses. To address these limitations, this study proposes a novel framework for embodied intelligent assembly robots, leveraging a synergistic collaboration between large-scale model (LSM) and small-scale models (SSMs). The framework integrates four functional layers - perception, decision, actuation, and feedback - enabling adaptive, closed-loop control. The perception layer fuses multimodal sensor data using LSM for semantic understanding and SSMs for real-time preprocessing. The decision-making layer combines LSM-driven global planning with SSMs-based high-frequency local responses. The actuation layer translates decisions into precise actions, while the feedback layer ensures dynamic optimization through hierarchical error correction. Key enabling technologies include real-time multimodal data fusion, hierarchical reasoning, collaborative instruction generation, and feedback-driven precision control. A proof-of-concept implementation was conducted on a simplified satellite prototype to verify the effectiveness of the proposed framework in the battery module reinstallation task. With manual assistance, the assembly success rate can reach 80%. These preliminary results validate the feasibility of the LSM-SSMs collaborative framework and demonstrate its potential to support high-precision, adaptive, and safe robotic assembly in aerospace manufacturing and related industries.