<p>Coaxial monitoring of the melt pool, the molten region formed during laser exposure in laser powder bed fusion (LPBF) additive manufacturing, is critical for ensuring part quality, as its morphology reflects process stability. However, melt pool images are inherently stochastic due to variations in processing conditions, imaging setups, and transient phenomena such as spatter. The impact of this variability on segmentation accuracy has not been thoroughly explored in the literature, resulting in a critical gap in the development of robust methods. As a result, existing segmentation techniques often lack the generalizability and reliability needed to perform consistently across diverse conditions. In response, we propose RAMPSeg (Robust Adaptive Melt Pool Segmentation) algorithm, an intensity-agnostic and calibration-free edge-detection-based method that introduces three key innovations: (i) data-driven optimization of edge detection parameters, (ii) a quantitative segmentation review using an edge-to-area ratio to guide refinement, and (iii) an adaptive smoothing feedback loop. Unlike prior methods, RAMPSeg avoids arbitrarily selected parameters, over- or under-segmentation, and fixed heuristics. Instead, it achieves self-regulating segmentation that dynamically balances noise reduction and boundary preservation across diverse imaging conditions, enabling more generalizable melt pool analysis. We comprehensively evaluated its effectiveness using three distinct datasets encompassing diverse process conditions, materials, machines, and imaging systems. This comprehensive evaluation demonstrated consistently high segmentation accuracy (~ 90%) against expert-annotated ground truth, significantly outperforming conventional thresholding, Otsu’s method, and state-of-the-art zero-shot vision transformer models (SAM and CLIPSeg). RAMPSeg enables reliable melt pool segmentation under dynamic conditions, supporting accurate anomaly detection, process optimization, and quality assurance.</p>