PAEAR: Point Clouds Area Exploration and Active Recognition method driven by reinforcement learning for robotic welding

For welded components without CAD models or prior geometry, robots face significant challenges in translating and interpreting design information into welding-relevant knowledge. When facing unfamiliar workpieces, welding robots cannot autonomously understand seam, edge, and surface elements. Consequently, welding seam localization and trajectory planning still rely heavily on manual teaching and experience-based search. In this study, a weld local feature descriptor (W-LFD) is developed to capture the local patterns of typical weld joints using interpretable geometric statistics, including inter-cluster surface angle, local density coefficient of variation, and normalized dispersion ratio. Together with small-sample pre-training, a prior feature library is established by coupling the descriptor with its parameters. Moreover, weld area acquisition is formulated as an active exploration problem in point clouds space. A reinforcement learning-based point clouds exploration and active recognition (PAEAR) method is proposed with a translation–scaling multi-level action strategy. This design enables identification and segmentation of regional point clouds for different welding seam categories in part-level scenes. Experimental results show that, across 2 simulated and 2 real-world scenarios, PAEAR achieves an mIoU of 86.86% using only 4 labeled samples, improving upon mainstream semantic segmentation methods by up to 81.22%, while reducing the area recognition error rate to 5.4%. The proposed method can serve as a front-end perception module for welding systems, providing reliable regional point clouds inputs for downstream welding seam localization and welding trajectory planning, while reducing dependence on manual teaching and prior models.