Facing the new generation intelligent manufacturing, traditional manufacturing models are transitioning towards large-scale customized productions, improving the efficiency and flexibility of complex manufacturing processes. This is crucial for enhancing the stability and core competitiveness of the manufacturing industry, and human-robot collaboration systems are an important means to achieve this goal. At present, mainstream manufacturing human-robot collaboration systems are modeled for specific scenarios and actions, with poor scalability and flexibility, making it difficult to flexibly handle actions beyond the set. Therefore, this article proposes a new human-robot collaboration framework based on action recognition and multi-scale control, designs 27 basic gesture actions for motion control, and constructs a robot control instruction library containing 70 different semantics based on these actions. By integrating static gesture recognition, dynamic action process recognition, and You-Only-Look-Once V5 object recognition and positioning technology, accurate recognition of various control actions has been achieved. The recognition accuracy of 27 types of static control actions has reached 100%, and the dynamic action recognition accuracy of the gearbox assembly process based on lightweight MF-AE-NNOBJ has reached 90%. This provides new ideas for simplifying the complexity of human-robot collaboration problems, improving system accuracy, efficiency, and stability.

中文翻译：

---

面向新一代智能制造的基于多尺度控制和动作识别的人机协作框架


面对新一代智能制造，传统制造模式正在向大规模定制生产转型，提高复杂制造流程的效率和灵活性。这对于增强制造业的稳定性和核心竞争力至关重要，而人机协作系统是实现这一目标的重要手段。目前主流的制造人机协作系统都是针对特定的场景和动作进行建模，扩展性和灵活性较差，难以灵活处理超出设定的动作。因此，本文提出了一种基于动作识别和多尺度控制的新型人机协作框架，设计了27种用于运动控制的基本手势动作，并基于这些动作构建了包含70种不同语义的机器人控制指令库。通过集成静态手势识别、动态动作过程识别以及You-Only-Look-Once V5物体识别和定位技术，实现了各种控制动作的准确识别。 27类静态控制动作识别准确率达到100%，基于轻量化MF-AE-NNOBJ的变速箱装配工艺动态动作识别准确率达到90%。这为简化人机协作问题的复杂性、提高系统精度、效率和稳定性提供了新思路。