Vision-based real-time autonomous trajectory planning in robotic laser welding applications

In the current industrial landscape, robotic laser welding remains highly dependent on skilled operators, particularly during setup and path programming. Existing seam tracking systems typically operate as correction layers on predefined trajectories, limiting flexibility and requiring extensive pre-welding preparation. This paper presents a vision-based framework for online trajectory generation without a predefined welding path. A coaxial camera provides real-time visual feedback of the welding zone, processed by a deep-learning-based module to extract seam geometry. The estimated seam is incrementally converted into cartesian motion targets and streamed to an industrial robot controller for execution. A key contribution is the latency-aware integration of perception, incremental trajectory planning, and a proprietary industrial robot controller. Instead of using a fixed control frequency, the system generates target poses in an event-driven manner, based on visual prediction availability and robot execution state. On the tested hardware, a nominal 35 ms perception-to-execution latency is managed through adaptive synchronization and partially overlapping trajectory segments. Experimental validation on linear and curved stainless steel and aluminum alloys yielded prediction-to-seam deviations below (Formula presented) with respect to manually annotated seam centerlines, while robot execution fidelity was quantified through command-to-execution cross-track deviations. Continuous execution was maintained even under initial positional misalignment.