A person inside the laser cell — and the perimeter scanner arrives just in time.

The problem

The laser cell sees the perimeter. What happens inside cannot be seen from the perimeter.

The safety of a robotic laser welding cell crosses three realities that are rarely governed together:

1. What the cell itself sees — safety laser scanners, mats, interlocks. First layer, SIL/PL-certified. Essential, but tied to the perimeter and with blind zones inside.
2. What happens from above — the information that would help: someone slipping in through a badly closed maintenance hatch, someone entering through the part-loading area before the cell has finished its stop, a maintenance technician leaning over the beam zone.
3. How much real authority the plant has over the cell's logic — the cell arrives closed from the integrator, and adding your own criterion without touching the certified logic is non-trivial.

The result: most near-misses in laser cells happen at zones with hatches or in the transitions between modes (automatic ↔ teaching ↔ maintenance), and are documented after the fact in the shift leader's report. The cell's safety system works 99% of the time. That 1% is what you want to armour with a second, anticipatory layer owned by the plant.

How it fits into the IRIS system

iLEAN does not touch the cell's certified safety — it adds an anticipatory layer on top, in line with ISO 10218.

The problem with the laser cell is not that the safety system is bad: it is that its criterion is fixed by the integrator and lives on an island relative to the plant. iLEAN acts as the filler that closes that gap, without touching what is certified.

Edge sees the protected zone from above. Connect sends the order to the safety PLC. The agent cross-checks with the cell's mode and the maintenance calendar. The safety manager signs the rules.

The three iLEAN pieces applied to intrusion detection in a laser cell:

• Edge — a terminal with machine vision (CNN) on an overhead camera above the cell, in visible, NIR or thermal depending on the environment. It recognises the human silhouette in the protected zone and sends the stop order to the safety PLC by dry contact or safe bus. It works with no network. If the plant loses its WiFi, Edge keeps protecting the cell.
• Connect — captures the robot's state, the program it is running and the mode (automatic/teaching/maintenance). In teaching mode the rule changes automatically — a person in the zone is expected. Connect also picks up external notices (the shift leader's near-miss report, the integrator's instruction).
• Agent — cross-checks Edge events with the cell's mode and the maintenance calendar. If it detects a pattern (a specific hatch producing repeated intrusions, a shift change with several events), it proposes a rule change to the safety manager and leaves the dossier. The person validates the change.

See the full IRIS architecture →

Before and after

Perimeter safety alone vs. perimeter + iLEAN Vision layer

Aspect	Laser scanner + interlocks only	With iLEAN Vision as a second layer
Perspective	Perimeter-bound, tied to the integrator	Overhead, sees the whole protected zone
Blind zones inside the cell	Covered as per the original design	Overhead coverage of zones with hatches or loading
False positives from laser reflections	Frequent with classic auxiliary cameras	The CNN classifies the silhouette, ignores glare
Rule change by mode (auto/teaching)	Manual and outside the system	Automatic change based on the robot's mode
Governance by the plant	Limited to the integrator's parameters	Rules configurable by zone and mode
Near-miss dossier	Reconstructed by hand	Automatic, with video and the robot's mode at the event

Impact estimate

Impact estimate for your plant — to be validated with your numbers.

The block below is an estimate to be validated with the specific figures of your plant. We set it out so the committee has an order of magnitude; we refine it during the diagnostic.

• Plant with 2-6 laser welding cells (body panels, exhausts, chassis or other assemblies), known integrator, safety PLC already installed.
• Edge pilot with 1-2 overhead cameras on the critical cell + integration with the PLC. First expected value within a few weeks.
• Indicative payback between 4 and 9 months, depending on the frequency of documented near-misses and the average cost of the integrator's forced stops to review an incident.
• Expected reduction in near-misses at zones with hatches or mode transitions: ≥ 30% versus the measured baseline. A single avoided incident with sick leave pays for the pilot with plenty of margin.

And the prevention manager's reasonable doubt

"What if the CNN gives a false positive and stops the cell for no reason?" — the costs of a false positive (1-2 minutes of stopped cell) are orders below those of a false negative (an incident with sick leave), so the rule is calibrated towards prudence. Moreover, hallucination is a problem of free generation, not of anchored tasks: classifying what is seen in an image against a closed catalogue of classes (person/robot/part/trolley) is exactly the kind of task where the best models brought the error below 1.5% ^[1]. And the cell's primary SIL/PL safety layer remains intact as the net underneath.

^[1] OpenAI paper "Why Language Models Hallucinate", 2025 — on the reliability of AI in anchored tasks.