A pre-rolled cone is a high-volume, low-value consumable, and the defects that matter are cosmetic and structural — a contamination speck, a tip that did not form, a loose filter fold, an out-of-round base. The line has to feed, image, decide, reject, verify and pack with no one in the loop, at a rate a person could not sustain.
Pre-rolled cones are produced in enormous volume and were inspected, historically, by eye. Manual inspection is inconsistent and does not scale, and the defects are small: a dark contamination speck on a pale paper body, a tip that did not form cleanly, a filter crutch whose rolled fold is loose or irregular, and a base opening that is not round.
The economics invert the usual inspection trade. The unit is cheap and the volume is high, so the inspection cost per unit has to be almost nothing — which means throughput, not per-image cleverness, is the constraint. The line inspects several cones in a single image so the cost per unit stays near zero, and it runs at many cones a second.
And it had to be a line, not a camera. Feeding, conveying, imaging, rejecting, verifying and packing all had to run as one automated flow, because an inspection step that produces a verdict but leaves a person to act on it saves nothing at this volume.
The ones that decide whether the line pays for itself, stated before any of them had an answer.
A sub-millimetre speck decides a reject on a low-value, high-volume unit. Imaging has to resolve the defect while the inspection cost per unit stays close to zero.
At many cones a second the whole line — capture, inference, decision and actuation — has to hold the rate, not just the model. The beat is shared.
Cones are light, tapered and nest together. Singulating them reliably from a stack without jams is harder than inspecting them, and it sets the line rate.
The vision result has to reach a pneumatic reject at the right cone, over PLC IO, before that cone has moved past the bin. A correct verdict applied late is a defect shipped.
The imaging and inference blocks are marked because the line’s economics live there — the defect has to be visible and the verdict has to land in a fraction of a second. The verification block is marked because it is a second, independent check that a rejected cone actually left the line.
The scope of the build, from feasibility to an installed line, rather than a capability list.
Rarely the subsystem that sounds difficult. Written out because a reader facing the same line gets more from this than from a list of what went well.
Singulating light, tapered, nesting cones from a stack without jams took more iteration than the vision did. The feeder sets the line rate, so its availability — not its speed — is the number that matters.
A fast per-image number is only useful if capture, transfer, inference, decision and actuation all fit the same beat. The tail of each stage, not its average, is what drops cones.
Resolving a sub-millimetre speck while keeping per-unit cost near zero drove the imaging, and the decision to inspect several cones in one frame rather than one at a time.
A verdict is only useful if it reaches the pneumatic actuator at the right cone. Tracking each cone from camera to reject bin had to be exact at line speed.
An independent scan after the sort verifies a rejected cone actually left, because a reject that does not physically drop is a defect that ships. The check is deliberately independent of the camera that made the verdict.
Feed, convey, inspect, sort, verify and pack running as one automated flow, from a manually loaded stack to a packed box.
Several cones per image on a single GPU — a design sized to run at many cones a second rather than to a benchmark.
Contamination specks, tip formation, filter-fold quality and base roundness — the classes manual inspection caught only inconsistently.
An independent scan confirms the sort, so the line’s promise is that a rejected cone has actually left it rather than merely been flagged.
Customer projects are presented at property, capability, outcome and integration level. Customer names, internal architecture, confidential deliverables and commercial terms are not disclosed. Where a figure would identify a customer, it is omitted rather than approximated. More detail is available under a non-disclosure agreement, within the limits the customer has agreed.
Every item links to its own page, with characteristics, applications and maturity stated honestly for that item.
The inspection solution, imaging to sort.
SOLUTIONFeeding, sorting and packaging.
PLATFORMThe machine vision platform.
PLATFORMControllers, IO and the device cloud.
CASE STUDYInspection on a different line.
KNOWLEDGEHow the handshake works.
Because the part is cheap and the volume is high, so the inspection cost per unit has to be almost nothing. Imaging several cones per frame and processing the frame in a fraction of a second on a single GPU is what makes the per-unit cost small enough, at many cones a second.
Cosmetic and structural classes: contamination specks on the pale paper body, tips that did not form cleanly, loose or irregular filter folds, and out-of-round base openings — the defects manual inspection caught only inconsistently.
As an independent check that a rejected cone actually left the line. A pneumatic reject that does not physically drop the cone would otherwise let a known defect through, so an independent scan verifies the sort happened rather than assuming it.
Feeding, not inspection. Singulating light, tapered, nesting cones from a stack without jams set the line rate and took more iteration than the vision model did.
Written for engineers. Share it with one.
Tell us the specification, the constraint and the deadline. Programmes that cross silicon, radio, embedded and AI are where Faststream is strongest.