AGV proximity at regulated warehouse intersections — zero near-miss and ISO 3691-4 traceability.

The problem

The intersection is the fleet's blind zone — and it is where ISO 3691-4 squeezes you.

An AGV is safe looking forward. Your problem is that the intersection is made of three things that come from the side:

1. Another fleet, or your own fleet, on the perpendicular. Each AGV decides its trajectory with the manufacturer's supervisor; when two supervisors do not talk to each other (because they come from different vendors, or because one fleet is subcontracted and the other is yours), the crossing is resolved defensively: everyone gives way and nobody moves.
2. The driver-operated forklift entering from the picking aisle. The AGV's scanner sees it metres away — and those are seconds too late to stop the AGV from braking hard with the risk of dropping its load.
3. The operator on foot stepping out of the racking aisle. The most vulnerable one. Floor markings and convex mirrors are what the audit asks for, but they do not see trajectory — only instantaneous position.

When the ISO 3691-4 auditor reviews the incident log and asks about near-misses, today you answer with whatever the operator wrote down by hand. That answer gets harder to defend every year — and every undocumented near-miss is the prelude to the one that does end in an injury.

How it fits the IRIS system

iLEAN does not replace the AGV's laser — it adds the overhead view the crossing was missing.

The intersection problem is not a lack of safety: it is a lack of cross-view. The certified functional safety layer (each AGV's laser scanner, safe bus, emergency stop) stays where it is. Fleet management stays in the manufacturer's supervisor. iLEAN Edge sits above the crossing, like the filler that covers what no individual system could see.

Edge sees the intersection from above. The agent cross-checks trajectories and warns the fleet supervisor before the laser has to fire. The warehouse manager has the per-event dossier for ISO 3691-4.

The three iLEAN pieces applied to the intersection:

• Edge — a terminal with an overhead camera above the crossing. The CNN detects classes (own AGV, external AGV, forklift, operator, pallet on the floor) and estimates trajectory in milliseconds. It works without network — if the warehouse WiFi drops, it keeps warning the local fleet supervisor over dry contact.
• Connect — captures the task plan from the WMS and the fleet supervisor, and absorbs what arrives from outside: an extra shift from the logistics provider that brings in an external AGV, a dock stop planned by the customer, a route change.
• Agent — consolidates every proximity event (which actors, what speed, what the system decided) into a dossier per intersection and per shift, useful for ISO 3691-4, for the insurer and for the health and safety manager. The person validates the policy; the agent applies it consistently 24/7.

See the full IRIS architecture →

Before and after

Intersection with laser + mirrors vs. intersection with the iLEAN overhead view

Aspect	AGV laser + mirrors + floor markings only	With iLEAN Edge + Connect + Agent
View of the crossing	What each AGV sees ahead of it	Overhead, all actors at once
Cross-trajectory	n/a — instantaneous position only	Estimated in milliseconds, by class
Mixed fleets	Each supervisor decides defensively	Agent coordinates at crossing level
ISO 3691-4 traceability	Logbook + verbal operator report	Per-event dossier with video-event
Near-misses	The ones somebody remembers to note	All of them, counted and classified
Certified functional safety	Laser scanner + certified bus	Untouched — Edge sits above it

Impact estimate

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

The block below is an estimate to be validated with your warehouse's actual data. We put it forward so the committee has an order of magnitude; we refine it during the diagnostic.

• Regulated warehouse in distribution, retail or production with its own AGV fleet (8-20 vehicles), 2-4 critical intersections where AGVs, forklifts and operators mix.
• Edge pilot on the Pareto intersection (overhead camera + integration with the fleet supervisor + ISO 3691-4 dossier). First value expected within a few weeks.
• Expected reduction in documented near-misses of ≥ 30% against baseline — and the rest recorded instead of invisible.
• Indicative payback between 4 and 9 months, depending on the historical frequency of near-misses and the avoided cost of an incident with injury.
• Hard lever: per-event traceability takes the ISO 3691-4 audit headache off your desk, and the throughput the fleet recovers at the crossing pays for the pilot on its own.

And the warehouse manager's reasonable doubt

"What if the AI mistakes an actor at the crossing and sends the wrong order?" — hallucination is a problem of free generation, not of anchored tasks. In tasks where the AI is limited to classifying visible actors and estimating trajectory (what can be seen, in 2D overhead geometry), the best models brought error below 1.5% ^[1]. And even so, the critical decision is never taken alone: the AGV's certified functional safety layer (laser scanner, safe bus, emergency stop) stays active and triggers the stop by itself if Edge fails. The three rings are there for exactly this.

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