Industry

Stop asking inspectors to outstare fatigue: cobot AI cell cuts QC from 82s to 61s and viewing time 82%

A Dongguan kitchen-appliance final-control line asks a senior inspector to stare down glass and sheet-metal surfaces across three shifts—one miss returns a whole lot. The board says “add people.” The cost sheet says “cannot.” Eyes are not infinite.

Ünal, Okasha, Çırakman, and Ünal (TEKNOPAR, AI-PRISM) turn that shop tension into checkable field results: at Silverline’s kitchen-appliance plant in Turkey, a collaborative inspection cell—UR 10e (vision defects) + Comau Racer-5 (functional tests), coordinated on ROS 2 Humble—cuts per-unit QC from 82 s to 61 s (25%), lifts final-control resource efficiency from 0.75 to 0.88 (+17%), trims station operators 3→1, cuts operator viewing time −82%, and significantly lowers NASA-TLX mental demand (p=0.005). This note maps plant management moves. It is not a reprint.

What you can verify in the paper

Problem: manual visual inspection remains a line bottleneck; long-shift fatigue lowers defect detection; lab algorithm scores are not plant clearance.

Method and results (wording stays with the paper):

• Architecture: UR 10e with Basler for defect scanning + Comau Racer-5 for functional tests; multi-modal data from Basler imagery, TIA acoustics, SPS electrical safety; Ubuntu 22.04 + ROS 2 Humble; Grafana live dashboards.
• Cycle time and efficiency: per-unit QC 82 s → 61 s (25%, beating the original 10% target); final-control resource efficiency 0.75 → 0.88 (+17%, target +15%), with operators at the station 3 → 1.
• Load: operator visual-inspection viewing time −82%; NASA-TLX mental demand significantly reduced (p=0.005).
• Perception: hood tracking accuracy ≈100%; 2D defect detection ≈85% (false positives on the white-hood variant dominate residual error); per-frame inference ≈0.25 s; end-to-end perception ≈0.5 s.
• Human-factor trade-off (stated plainly): occupational self-efficacy 5.13 → 4.95; technology readiness 3.18 → 2.88—economic and load gains arrive with a modest acceptance dip that needs continued training and engagement design.
• Deployment challenges: close-proximity safety, glare/reflections, ROS 2 namespace collisions across two cobots, OS/dependency issues; HRI structured via Rodríguez-Guerra’s four levels (task / operation / work cell / process).

Authors’ takeaway: collaborative inspection should augment rather than replace the operator—automate scanning, keep judgment human; most deployment cost sits in integration and lighting, not model scores alone.

Source of record: arXiv:2609.33522 (PDF). Title: AI-Driven Collaborative Assembly Line Inspection: System Integration and Deployment Challenges.

Pearl River Delta translation: redesign the QC cell, do not only buy a camera

Most shops do not need a higher-scoring defect model in isolation. They need a staffing rule that hands repetitive scanning to machines, keeps accept/reject with people, and writes glare and safety into the cell standard. A workable minimum:

1. Change task allocation before chasing headcount math — the paper’s 3→1 shift assumes every flag is still human-reviewed; without that loop, ≈85% defect accuracy cannot carry final-control liability.
2. Put glare and SKU variants into acceptance — white-hood false positives dominate residual error; pilot gates should bucket by color/material, not report one headline accuracy.
3. Isolate namespaces before dual-arm go-live — ROS 2 topic collisions were a field blocker; any two-cobot co-line plan needs namespace isolation and launch orchestration before production.
4. Budget training when technology readiness dips — modest drops in self-efficacy and readiness are a deployment trade-off, not a single “system is bad” verdict; engagement and decision rights must ship with the hardware.

Organization power here means locking QC cycle time, station staffing, cell safety, and acceptance into one pilot scorecard—not buying another isolated “99% accuracy” black-box camera.

A fifteen-minute self-check

Was any final-control miss this week “solved” only by asking someone to stare harder—with no discussion of automated scanning plus human review? If vision alerts are switched off, can the shift lead re-staff viewing within five minutes under station red lines? Two blanks means “outstare fatigue” is still the operating theory.

Public wording stays with the arXiv paper. This note only maps plant management moves.