Scalable Shared Encoding Architecture for Learning-Based Error Detection in Robotic Wiring Harness Assembly

This paper focuses on an automatic solution for the detection of manufacturing errors in the context of automatic wiring harness assembly. In the proposed setup, a robot grasp the wires and places them in specific assembly clips according to the wiring harness design. However, due to the deformability of cables, the process outcome is not completely predictable since sometimes the cables remains entangled in other parts of the assembly or do not fit properly into the clips. The proposed error detector verifies the correct insertion of each cables within the clip, considering that the number of cables and their dimension change along the different assembly stage. The proposed solution covers possible state-of-the-art network learning model that use point clouds as input source, while the network architecture is designed to offer precision and scalability in the context of a flexible and dynamic automation. The developed solution achieved a 96% precision on a dataset composed by various scenario. Therefore, despite being conceived for a robotic wiring harness manufacturing system, the proposed solution can be potentially applied as an online quality control system in manual wiring harness manufacturing.