Leveraging CNN and LSTM-based perception towards reconfigurable human-robot collaborative operations
提出一种结合CNN和LSTM的动态感知方法，使工业工作站能自动适应产品变化和操作员状态，减少生产延迟并保持高装配可靠性。
Efficient human-robot collaboration requires the ability of industrial workstations to be easily reconfigured. However, conventional systems rely on predefined task execution and limited perception capabilities, restricting their ability to adapt to product variability. This paper presents the implementation of a human-robot collaborative paradigm for reconfigurable industrial operations utilising dynamic perception capabilities. The proposed approach follows an experimental evaluation methodology based on a production time cost model at workstation level. Process and human perception techniques are deployed to dynamically adapt to changes in product variations and environmental scene updates. A Convolutional Neural Network (CNN) based pipeline is introduced for automatically retraining and accurately detecting diverse industrial parts. A Long Short-Term Memory (LSTM) based module is also deployed to identify the operator’s state and adapt robot behaviour. The synchronisation of execution flow is achieved through a service-based orchestration architecture acting as a state holder for monitoring the overall system while incorporating results provided by the perception modules. The system is validated through a case study derived from the white goods industry, using key operational metrics including production time, system utilisation, and assembly reliability. Experimental results demonstrated improved workstation adaptability and reduced operational delays when new components are introduced while maintaining high assembly reliability.