VISION-BASED DRY-MATERIAL INVENTORY ESTIMATION IN AUTONOMOUS LABORATORIES: RGB SEGMENTATION VERSUS RGB-D SURFACE RECONSTRUCTION

Abstract

Reliable dry-material inventory monitoring is needed for autonomous laboratory operation, but the suitability of a non-contact method depends on material behaviour, optical access, calibration requirements and the quantity to be estimated. We compared two vision-based approaches—material-specific RGB/YOLO level estimation and visible-surface reconstruction from an Intel RealSense D405—across three materials (black polypropylene pellets, wood powder, and microcrystalline cellulose fine powder) in small transparent laboratory tubes representative of common dosing instruments, with the aim of informing a practical, evidence-based framework for choosing between such methods in laboratory automation and similar industrial contexts. Software-coordinated records linked the two camera channels to a nominal mass-fraction inventory reference and a manual-height baseline. Across six runs and 133 paired, reference-eligible material states, D405 achieved lower error than YOLO in every run (mean absolute error of 3.26 vs. 7.14 percentage points), though YOLO offered higher valid-result coverage (97.2% vs. 95.1%). Both methods showed increased error under repeated surface disturbance, with YOLO degrading more sharply than D405. An exploratory correction fitted using the first observation at each mass level substantially improved YOLO but did not surpass D405’s uncorrected accuracy, while a single global correction worsened both methods—indicating that simple corrections do not transfer robustly across conditions. The results support a method-selection framework based on error, availability, viewing geometry, calibration-state representativeness and total resource requirements. Conclusions are limited to the tested materials and setups, and the mass-fraction reference is an operational inventory proxy rather than a direct measurement of true occupied volume.