> ## Content Index
> Fetch the complete content index at: https://blog.robovision.ai/llms.txt
> Use this file to discover other available public pages before exploring further.

# You're Cutting In the Right Place. You Just Don't Know Exactly Where That Is.
- URL: https://blog.robovision.ai/lettuce-decoring-extractacore-qing/
- Published: 2026-09-22T10:03:41.000Z
- Updated: 2026-10-01T10:50:07.000Z
- Description: Automated lettuce de-coring has accepted 30 to 40% waste for years. Extractacore and QING cut that to 3 to 5%, without changing the machine.
- Author: Robovision AI
- Tags: food processing, vegetable processing, machine vision, OEM

Your customers grade a new crop, a new animal, a new bake every week, and expect the machine to keep up. Add that capability without building the infrastructure yourself. You keep the result. Governed for the life of the model, delivered with our team.

In fresh produce processing, the margin between a good product and a wasted one is sometimes measured in millimetres. Nowhere is that more literal than automated core removal on a lettuce line. Every head needs its core cut out cleanly: not so shallow that residual core remains, not so deep that usable product is sacrificed, at industrial speed, continuously, across a full shift. For years, the industry answer to that problem was a wide margin of error dressed up as a specification.

## What changed to make 30 to 40% de-coring waste a competitive problem?

For years, 30 to 40% de-coring waste was the accepted cost of reliable operation, not a number anyone expected to move. That changed once a governed 3D vision upgrade proved a line could hold 3 to 5% waste in production, at full speed, without touching the mechanical system.

Once one processor has seen that number, every OEM still quoting the old one is defending a status quo its own customers can now measure against a real alternative. That shift matters more than a single case study. Processors compare machines on throughput and uptime as a matter of course. A yield number that used to be an assumed constant, buried in the cost of doing business, is now a line item a buyer can ask an OEM to defend. The conversation about de-coring equipment has permanently changed shape.

## Why do de-coring machines cut wider than the core actually is?

Because for years, no vision system could tell the cutting tool exactly where the core was, so the machine compensated with a mechanical safety margin. Removing more than necessary guaranteed no residual core ever reached the next station. That reliability came at a fixed, repeating cost: a meaningful share of every head, sacrificed as usable product on every single cut.

The limitation was never the robot or the cutting tool. Both were capable of cutting exactly where instructed. What didn't exist was a system that could reliably tell them where that was: the exact position, angle and depth of a core that shifts from head to head. Absent that, wide-margin cutting was the only strategy that held up at one head per second, across a full shift, without a residual-core escape. The yield loss was not a defect in the system. It was the system working as designed, at a cost nobody had the data to challenge.

## Why can't a conventional 2D vision system fix it?

Because a 2D system matches patterns against training data, and falls back on an average when the match is weak. Lettuce varies by head, batch, season and days since harvest, so real production conditions constantly drift from whatever the system was calibrated on. The system cannot reason about that variation. It can only manage it, and cutting wide is how it manages it.

A conventional vision system captures an image, applies rules calibrated at commissioning, and outputs a position estimate. When the incoming head resembles the training data closely, the estimate holds. When it doesn't, and in fresh produce it constantly doesn't, the system falls back on an assumption rather than a measurement. That gap between commissioning conditions and week-forty conditions is not a rare edge case. It is the normal operating state of a line handling a biological product, and it is the real defect a de-coring line has to solve: not a missed core, but an unreasoned position estimate standing in for one.

## Where does detection have to happen, and what does it take?

Detection has to happen before the cut, on every single head, at line speed, producing a measurement instead of an estimate. QING integrated Robovision's Indurion Suite into Extractacore's existing lettuce line to do exactly that, mapping each core's position, depth and orientation in three dimensions.

It is the same class of work Defect Measurement does elsewhere in the Suite, segmenting and sizing irregular geometry for objective measurement, here pointed at a core boundary instead of a defect: the output is a precise coordinate set and cutting angle for the robot, per head, instead of a fixed margin.

|                                    | Fixed-margin 2D vision                | Governed 3D AI vision                           |
| ---------------------------------- | ------------------------------------- | ----------------------------------------------- |
| Core localization                  | Pattern match against training data   | Measured position, depth, orientation, per head |
| Cutting margin                     | Fixed, sized for worst-case variation | Set by the actual geometry of that head         |
| Response to head-to-head variation | Falls back on averages                | Reasons about each head individually            |
| Mechanical changes required        | \-                                    | None: robot, conveyor, cutting tool unchanged   |
| Waste, in production               | 30 to 40%                             | 3 to 5%                                         |

The robot, conveyor and cutting tool at Extractacore's line did not change. What changed was the intelligence guiding them: a governed decision layer that turns a naturally variable product into a set of individual, measured cutting instructions instead of one fixed rule applied to every head.

## From 30 to 40% waste to 3 to 5%: the result in production

The system now processes approximately one head per second, roughly 55 crops a minute. For every head, the vision model produces an X, Y, Z coordinate set and cutting angle from that head's own geometry, not an estimate derived from averaging. The robot cuts exactly there. Waste dropped from 30 to 40% to 3 to 5%, measured in production, not projected.

*"To guarantee full core removal, machines tended to cut wider than necessary. That ensured safety and consistency, but it also meant losing usable product."* (Letti Barber, Founder & CEO, Extractacore)

*"Variability isn't the enemy. Ignoring it is. AI allows you to work with natural products on their own terms."* (Teun Keusters, Product Manager, QING)

For Extractacore, the commercial implication is direct. The cutting mechanism already installed at its customers' sites now performs at a fundamentally different level, without replacement or major modification. The upgrade path was the vision layer alone.

## Why is this a governance problem, not a one-time calibration?

Because lettuce keeps varying after installation, the way every biological product does: by season, by batch, by days since harvest. A model validated once and left alone drifts the same way a fixed mechanical margin does. What holds 3 to 5% waste through a full season is a model that keeps learning from logged production data, not a calibration frozen at commissioning.

That is a governance question, not a modeling one: versioned deployments, a rollback path if a retrain goes wrong, and a record of what the model saw and why it cut where it cut. A one-time recalibration is structurally doomed against a product that never stops changing. A governed lifecycle is the only architecture built to survive the schedule lettuce sets on its own.

## Does the same approach work on other vegetables besides lettuce?

The AI model continuously collects data on product characteristics and cutting performance, so the system improves with production instead of degrading from it. The same underlying capability, 3D AI vision, model governance, edge inference and continuous retraining, is directly transferable to other core-removal and precision-cutting problems: broccoli, cauliflower, cabbage, and beyond.

A single investment in the intelligence layer becomes the foundation for an expanded product line, not a one-off upgrade for a single crop. For Extractacore, and for any OEM or system integrator evaluating where AI fits in their equipment, this case points to something specific: the value of the machine is increasingly set by the precision of the intelligence guiding it. Where that intelligence is limited, yield loss fills the gap on every shift. Where it is governed and kept current, the gap closes and stays closed.