Unmanned AI FFB Grading
The grading station arrives
as a container.
Craned onto a plinth beside your ramp, wired to the weighbridge you already have. Four channels read every bunch. No grader needed.
“World’s first palm oil mill to adopt AI” — The Star, 19 Nov 2024 · See the rig on real fruit
The rig’s own output
This is the rig looking
at real fruit.
Not a diagram. A frame lifted straight out of the grading system’s own output on a live belt, the machine’s own masks and class labels burnt in. Everything coloured was drawn by the rig; the brackets and ledger below are ours.
- 01–03 The three regions the grading system localised in this frame. Each bracket sits on the position listed below, and there is no fourth.
- FILL Inside a bracket the rig’s own segmentation fill runs at full brightness; everything outside it is dimmed.On paper the wash and the reveal are dropped, so the frame prints undimmed end to end; the brackets, labels and coordinates carry over. Nothing is added, moved, sharpened or recoloured.
- RGB The colour channel is one of four. Socket, spikelet and thermal read the same bunch on the tiers that follow.
- Source file
- assets/img/vision-overlay-redacted.jpg
- Frame
- 1280 × 720 · grading rig, FFB conveyor
- Zone 01
- x549 y352 → x744 y460
- Zone 02
- x704 y480 → x921 y629
- Zone 03
- x450 y563 → x663 y706
- Overlay origin
- the grading system, not this page
- Withheld
- Confidence figures — no stated method or sample. Three regions localised; coordinates as shown.
The machine drew these masks and labels at grading time. Each label and its confidence score is blurred, because those figures come with no reference method, no sample and no approver. Nothing else is altered; the blur is confined to those three strips.
The bracket edges are soft, so the coordinates above are good to about ±10 pixels. They record where the machine drew on this frame — checkable by anyone who opens it — and say nothing about how well the system grades. The pale …FFB-GRADING lettering at the lower right is the grading system’s own on-screen title, burnt into its own recording.
Where the load comes from
A mill’s intake is other people’s fruit. Every load is a supplier who will want to know how the grade was reached.
Where it sits
Beside the ramp, on its own plinth, wired into the weighbridge.
How it works
What is being read
Ripeness is not one signal: the colour of the fruitlets, how many have detached, the sockets where they left, and the fibre between them.
Four channels read the
same bunch.
The rig takes one bunch at a time and turns it so every side is seen. Four tiers then read it. Each asks a different question about the same fruit, and each adds to the one before it.
TIER 0
Individualizer & rotator
Separates the bunch from the load and turns it through 360°, so the cameras see every face, not whichever one landed upward.
TIER 1
RGB colour
Colour for ripeness — the same signal a grader uses on the ramp, but taken at a fixed distance under a fixed light, on every bunch instead of a sample.
TIER 2
Socket analysis
The sockets left by fruit that has already dropped. Colour can be argued with. Sockets cannot.
TIER 3
Spikelet detection
Spikelet detail shows freshness — how long ago the bunch was cut. That is a different question from how ripe it was.
TIER 4
Thermal imaging
Heat, not surface. A bunch cannot be dressed up to fool a thermal camera.
As printed in the AIREI AI Grading Brochure, page 2 (The Unique Multi Tier Grading).
Lorry to ramp
On the belt
The rig does not stop the line. Bunches arrive at whatever rate the ramp feeds, and grading keeps up or it is a laboratory, not a mill.
What happens to a load.
- FFB is unloaded by the supplier.
- The auto feeder brings the fruit into the grading box.
- Inside the grading box, each bunch is individualised and rotated.
- Four-tier inspection: RGB colour, socket, spikelet, thermal.
- Automatic sorting on the AI result.
- Loose fruit is extracted separately.
- Ripe fruit goes to the loading ramp.
- Rejected fruit is returned to the supplier.
At the gate
Rejects go back on the lorry they came in on, and the chit is issued at the point of grading — a supplier leaves with the result, not a promise of one.
The eight steps are those of the AIREI corporate film Unmanned AI FFB Grading, in the film’s own order. Sent on request.
What the mill sees
The grade stops being
an argument.
The commercial value is not the camera. It is that grade, weight and supplier become one record, made at the same moment, that nobody in the queue had to agree to.
The console
Grading moves off the ramp, out of the heat and dust, onto a screen with a record behind it.
01
Control room
The grading line is watched from a room, not from the ramp. The cabin acknowledges; it does not assign the grade.
02
Long-stalk detection
Long stalk is reported as its own percentage, alongside freshness. The deduction suppliers dispute most becomes a measurement, not a judgement call.
03
Supplier statistics and historical pattern
Quality accumulates per supplier. Their record becomes a trend you can show them, and a basis for pricing that survives a conversation.
04
Weighbridge integration
Grade and weight are captured against the same vehicle in the same transaction, so the two records cannot drift apart.
05
Instant digital grading chit
The e-grading chit is issued at the point of grading. Nobody types it in afterwards.
Where the two numbers meet
Suppliers argue about the deduction, not the weight. Put grade and weight on one ticket and there is one document to talk about.
As printed in the AIREI AI Grading Brochure, page 2 (Key Features).
Accuracy, and what backs it
The instrument
A camera, a light bar, a frame, and a mill's own dust and vibration. Accuracy is a property of that whole assembly, not of the model inside it.
Ask for the accuracy figure,
and what comes with it.
What a site visit confirms today. Every bunch that enters the grading box is read on four channels: RGB colour, socket, spikelet and thermal. The machine issues the grade; nobody types it in. Stand on the platform and watch it happen.
What a visit cannot tell you is whether that grade is right. There is an accuracy percentage in AIREI’s own brochure. It goes in the box beside this when five things travel with it: the reference it was graded against, the sample size, who ran the comparison, which model build produced it, and who signed it off. Ask, and we send it with that basis.
A grading figure with no reference method is not a specification — it is a sentence.
Throughput works the same way. Ask, and we will send the figure the commissioning record supports.
Reserved
Grading accuracy, measured
- Accuracy figure
- Available on request. We send it with the reference method it was compared against, the number of bunches and the period, who ran the comparison, and the build of the grading model that produced it.
- Why it is not printed here
- An accuracy number without those four things beside it is not checkable, and we will not print one that is not.
- Reviewed
- First review 23 Aug 2026 · next 22 Feb 2027. Every later review is added to this line, so a date that moves shows as a date that moved.
The review date is not a release date. On 22 February 2027 this slot publishes either the figure or the row still missing.
Who signs it off
The figure ships when its reference, sample, tester, model build and approver ship with it.
Next step
Two ways in, and neither
of them is a brochure.
A site assessment tells you whether the rig fits your ramp, your weighbridge and your queue. The spec is what your engineer wants first.
What a visit settles
Ramp geometry, weighbridge wiring, queue length, and what the retrofit has to work around.
Or talk to an engineer now // WhatsApp ↗ // +603-8685 0888