A novel robotic multi-layer multi-pass welding adaptive path generation method based on point cloud slicing and Transformer for saddle-shaped weld seams

Research question

How can multiple pass trajectories and locally appropriate welding speeds be generated together for a saddle-shaped groove?

Method

Region growing and RANSAC establish pipe axes and the slicing reference for groove-feature reconstruction. A layer-filling strategy plans the paths, while a decoder-free Transformer learns the relation between slice sequences and welding speed.

Main findings

For the reported specimen, the complete computation for 25 passes took 21.917 s, averaging 0.877 s per pass. Path RMSE was within 0.5 mm and speed-prediction RMSE within 0.7 cm/min, followed by multi-layer, multi-pass welding experiments.