Can an AI-guided robot inspect hidden holes? Thirty trials stayed below 0.065mm position error
A proof-of-concept system combined adaptive lighting, AI segmentation and robot calibration. Every trial detected its target, but the experiment was small, unreplicated and partly confounded.
The short version
- What happened
- The peer-reviewed experiment tested 30 physical holes spanning 5–15mm diameters, 10–45mm depths, three reflectivity conditions and approach angles from 0° to 20°.
- Who it affects
- Manufacturing inspectors could spend less time repositioning parts if non-contact inspection proves reliable, but workers still need traceable calibration and authority to reject uncertain measurements.
- What remains unproven
- Thirty unreplicated trials in a partly crossed design
Keep in mind
Thirty unreplicated trials in a partly crossed design
No fixed-lighting, non-AI or alternative-system comparator
Reflectivity, angle, depth and diameter were partly confounded
No production contamination, burrs, oxidation, roughness or very deep holes
One system, one laboratory and one author
Source analysis · 1 linked source. Primary analysed source date: 26 October 2026.
Read the full report · 3 min The evidence, methods, implications and limitations
The robot measured every test hole within the study's narrow range
A peer-reviewed Indian engineering study tested a non-contact system for inspecting blind holes—openings whose base is difficult to see or reach. A six-axis robot positioned a camera and adaptive light; an AI model segmented the rim; sub-pixel fitting estimated its geometry; and calibration against a coordinate-measuring machine corrected robot-camera error. The 30 held-out tests covered 5–15mm diameters, 10–45mm depths, three surface-reflectivity classes and approach angles from 0° to 20°.
Every target was detected. Mean three-dimensional position error was 0.0322mm, with a 0.0645mm maximum. Mean absolute diameter error was 0.0544mm and average repeatability was 0.0217mm. The images used for these measurements were separate from the segmentation model's training and validation images, reducing direct test leakage.[1]
The result is a bench proof, not an autonomous factory inspection claim
The experiment gives engineers a concrete feasibility signal, particularly for aerospace, automotive and machined parts where moving a component to a separate measurement machine adds time. It does not establish reliability on production lines. The 30 conditions were not replicated, and the design partly tied reflectivity to approach angle and depth to diameter. The reported correlations therefore cannot identify which factor caused an error change.
There was also no matched fixed-lighting or conventional image-processing comparator, so the contribution of AI segmentation cannot be isolated from lighting, geometry fitting and calibration. Burrs, contaminants, oxidation, roughness and deeper holes were not tested. Confidence would increase with replicated factorial trials, traceable uncertainty budgets, ablation of each module and external validation on real industrial parts. The author reported no external funding or competing interests and released the dataset and code.[1]
The distinction that matters
What the 30-test proof of concept measured
The study compared the assembled system with coordinate-measuring-machine reference values, not with another inspection workflow.
Position
Robot, camera and calibration locate each hole
Mean error 0.0322mm; maximum 0.0645mm
Diameter
AI segmentation plus sub-pixel fitting estimates the rim
Mean absolute error 0.0544mm
Generalisability
Thirty unreplicated, partly confounded trials
Factory reliability remains untested