ActiveInspect: GRPO-Optimized Multi-Sensor Evidence Selection for Industrial Defect Detection

Automated visual inspection is a cornerstone of modern manufacturing quality assurance, yet the effectiveness of any detection system is fundamentally bounded by the informativeness of the observations it receives. Most vision-language model (VLM) and reinforcement learning methods for industrial defect detection assume a fixed set of observations and optimize only the reasoning applied to them. We introduce ActiveInspect, which formulates inspection as budget-constrained sequential selection of multi-view, multi-modal evidence. Starting from a pre-acquired observation pool, a single policy selects an additional view or modality, zooms into a candidate region, retrieves a matched normal reference, or terminates with a verdict. The policy is initialized by perception-activated supervised fine-tuning (PA-SFT) and subsequently optimized by group relative policy optimization (GRPO) using inspection-specific rewards. Depth and point-cloud measurements are converted into VLM-compatible geometric renderings, while a structured memory integrates evidence across inspection steps. Evaluation on Real-IAD D³, Real-IAD, MVTec 3D-AD, MVTec-AD, VisA, and MMAD demonstrates a consistent improvement in the accuracy-observation trade-off. On Real-IAD D³, ActiveInspect increases image-level area under the receiver operating characteristic curve (I-AUROC) from 0.890 to 0.906 (mean over three training seeds; p<0.01) relative to the passive D³M baseline while reducing the average observation count from 3.0 to 2.7. It reaches 99.8% of the I-AUROC obtained by exhaustive evaluation of all 15 observations while using 18% of that observation count, and it reduces per-sample inference time by a factor of 5.3 relative to the exhaustive scan. The largest gains occur for geometry-dependent defects, including dents, warping, and concavities.