A Review of Multimodal Perception and Closed-Loop Task Execution for Human-Robot Collaboration Robots Oriented Toward Intelligent Assembly

Abstract

With the increasing demand for flexible manufacturing, especially for multi-variety and small-batch production, human-robot collaboration has become more widely used in intelligent assembly. Many studies have investigated multimodal perception, task alloca tion, safety control, and quality inspection, and considerable progress has been made in these areas. Even so, several problems are still difficult to address. Information exchange between different technical modules is often not well connected, task under standing is still limited, and feedback from the assembly process is not fully incorporated into decision making. This review discusses recent research from the perspective of a complete "perception-decision-execution-feedback" process. The main topics inc lude operator action and intention recognition, perception of part states, task allocation under human-related constraints, safe trajectory planning, and quality feedback. Rather than considering these techniques separately, attention is given to how they interact during collaborative assembly. Current studies suggest that the main challenge is no longer improving the accuracy of a single perception or control module. A more important issue is how to connect multi-source perception, task understanding, safety constraints, and quality feedback into a stable closed-loop framework. Future work is expected to pay more attention to cross-station adaptation with limited training samples, practical and verifiable use of vision - language-action models, digital twin-based safety validation, and skill updating based on assembly deviations. It is hoped that this review can provide useful ideas for future research and system development in intelligent human-robot collaborative assembly.