AsperCut, the asparagus-harvesting robot: how 3D vision and AI cracked the detection challenge

2020. Covid grounds seasonal workers, and asparagus growers look out over their fields with no idea who will harvest them. Two engineers see a problem to solve. Four years later, their robot cuts, picks up and crates asparagus on its own. In between, they had to answer a question mechanical engineering could not handle: how can a machine see an asparagus spear?

KEY TAKEAWAYS IN 5 POINTS
- With 100,000 spears to harvest per person per season, manual harvesting is a major challenge for a French asparagus sector that is growing but under severe labour pressure.
- Sylektis first tried mechanical detection using feelers: not reliable enough, and liable to damage the product.
- 2D and 3D vision combined with AI removed the bottleneck, detecting each spear and giving the robot the exact coordinates of each harvesting point.
- As an added benefit, the data collected produces harvest maps that serve precision agriculture.
- The specific difficulty of this project — detecting a product that is different every time, in an environment that never stops changing — is exactly what any special-purpose machine builder faces when working outdoors.

Why asparagus harvesting is still a 100% manual operation

The paradox of French asparagus comes down to two figures: a sector that is gaining ground, and a harvesting method that has not changed in a century.

7,395 hectares in 2024: a growing French sector

The area under production is increasing. France had 7,395 hectares of asparagus in production in 2024, compared with 6,676 the previous year — almost 11% more land in a single year. Volumes, however, are flat: 27,476 tonnes harvested, versus 27,286 the year before, with average yield falling from 4.1 to 3.7 tonnes per hectare (source: Agreste, 2024 annual agricultural statistics, final figures, November 2025). In other words, every year there is more ground to cover for roughly the same tonnage.

Asparagus is a perennial, long-term crop: eight to ten years of production after planting. It also commands good prices. A grower who plants is committing for a decade.

400 to 500 labour hours per hectare

On the other side of the equation, an operation that has remained entirely manual. At its technical day in Castelnaudary in 2024, AOPn Asperges de France (the French national association of asparagus producer organisations) put harvesting at 400 to 500 hours per hectare, for a cost of around €1.50 per kilo. On top of that come 100 to 200 hours per hectare at the washing and packing station, plus one full-time equivalent for every 20 to 30 hectares for crop management alone — ridging, covers, irrigation (Christophe Paillaugue, President of AOPn).

The equation is simple, and that is exactly what makes it a dead end: the more the planted area grows, the more hands are needed. And hands are getting harder to find.

A European problem, not just a French one

This is not a French issue alone. In Germany, Europe's leading producer, the increase in the minimum wage to €13.90 per hour on 1 January 2026 was described as a major challenge by industry representatives, who are calling for a special scheme for seasonal workers.

When labour makes up most of the cost price and that cost is rising structurally, it is no longer a cash-flow issue. It is a business-model issue.

Sylektis: from a shortage of seasonal workers to the AsperCut robot

That observation is what gave birth to Sylektis. Wilfried Garrigue and Stéphane Abraham founded the start-up in the middle of the health crisis, after identifying the labour shortage as the breaking point for two crops in particular: asparagus and strawberries.
Four years later came one of the first open-field harvesting robots — first for white asparagus, then for green: the AsperCut.
In between, a detour through a technical dead end. That is the part of the story that matters to every special-purpose machine builder.

First attempt: why mechanical detection with feelers failed

Sylektis's initial reasoning was flawless. You can see a white asparagus spear when it breaks through the ridge. To keep costs down and offer an affordable machine, the team first developed mechanical detection: feelers that touch the spear to signal its presence.

Inexpensive. Simple. Robust on paper.
Out in the field, the solution did not hold up. Worse, it risked damaging the produce — something the growers flatly refused. On a crop where the spear is sold by the kilo, and where the crown determines the harvests of the next eight years, damaging it is not an option.
Six months of mechanical development led to one conclusion: the solution was not mechanical.

The team then turned to vision-based detection and approached APREX Solutions, a French designer of machine vision and image-based artificial intelligence solutions, based in Nancy.

The real challenge: detecting the unknown in an uncontrolled environment

To understand why this was no ordinary machine vision project, Sylektis's own words are hard to improve on:

“In industry, it's often about detecting something known in a controlled environment. Here, it's different every time, in a changing environment.”
Wilfried Garrigue – Co-Managing Director, Sylektis

That sentence says it all.

Machine vision on the shop floor vs. machine vision in the field

In conventional quality control, you inspect standardised parts, under controlled lighting, at a fixed distance, at a steady rate. You know what you are looking for, and you know what it looks like. The uncertainty lies in the defect, not in the object.

In the field, that foundation disappears. You do not know exactly what the object to detect looks like, nor under what conditions you will see it. The vision system has to decide quickly, without error, and never sees the same scene twice.

The variables that change with every spear

At least six parameters vary constantly:

1. The height of the spear, depending on its stage of maturity
2. The curvature and orientation of the stem, never perfectly vertical
3. The colour and texture of the soil, from one plot to another and from one region to another
4. The light, from dawn to low-angle full sun
5. The weather: rain, wind, dust, sand
6. Possible confusion with a stone, a clod of earth or plant debris

By the APREX teams' own admission, the project was well outside their usual scope. That is precisely what made it a good challenge.

How the vision system detects white and green asparagus

A starting point often overlooked outside the sector: white and green asparagus do not grow the same way. White asparagus grows underground; green asparagus grows in the open air. White spears are therefore harvested just as their tips emerge from the ridge, while green spears are cut above ground once they reach 20 to 30 cm.

Two growth patterns, two harvesting methods, two types of sensor. But one and the same detection building block, developed by APREX engineers — first for white asparagus, then extended to green.

White asparagus: combining a 2D camera and a 3D camera

For white asparagus, it was the combination of a 2D camera and a 3D camera that achieved the goal.

The challenge: detecting barely visible spear tips that can be mistaken for stones — or the other way round — and that blend into the colour of the soil depending on the terrain. Then sending their coordinates to the robot.

2D provides texture and colour information; 3D provides relief. On their own, neither is enough to make the call. Together, they remove the ambiguity.

Green asparagus: 3D point cloud and artificial intelligence

The solution developed for green asparagus — the case shown in the video — relies on 3D vision: a stereoscopic camera mounted at the front of the machine, above the harvesting plane, combined with artificial intelligence algorithms.

A stereoscopic camera works like a pair of eyes: two offset sensors observe the same scene, and the difference between the two images makes it possible to calculate a distance for every point. The result is no longer a photo, but a three-dimensional reconstruction of the plot.

Detect, measure, decide, guide: the 4 processing steps

1. The camera maps the ground. It produces a point cloud — a set of coordinates describing the real geometry of the scene: the soil, the shape of the ridge, and the spears emerging from it.
2. The AI interprets the point cloud. This is the core of APREX's development: an artificial intelligence technique able to work directly on 3D data to identify what, within that relief, is an asparagus spear — and what is not.
3. It extracts usable measurements. Tip coordinates, base coordinates, spear height. From these comes the decision: is this spear long enough to be harvested now?
4. It sends the harvesting coordinates to the robot. The robot cuts, grips, places the spear on a conveyor and fills the crates autonomously.

The four steps run continuously.

AsperCut in figures: harvesting robot performance
–    The AsperCut harvests 40 to 80 green spears per minute and up to 24 white spears per minute. By comparison, manual harvesting runs at 10 green spears and 6 white spears per minute.
– The AsperCut covers up to 4 hectares of white asparagus and up to 10 hectares of green asparagus. Each seasonal worker harvests 0.5 to 1 hectare by hand.
– Thanks to its hybrid power train, it can run 24/7.
–    100% made in France: designed and built in France, it carries a vision system developed by a French company that trains its own AI models, with data also hosted in France.

The AsperCut is supported by the French government's Investments for the Future Programme (PIA) and France 2030.

From harvest to data: vision serving precision agriculture

Vision-based detection brings an additional benefit: the value of the data.

A machine that sees also records. Every pass through the plot generates images and coordinates, stored securely. That material opens up a second use: harvest mapping. Where was the yield highest? Where does it drop off? And therefore: where should fertilisation, irrigation or maintenance be adjusted? This is precision agriculture in its fullest sense — not by adding a dedicated sensor, but by making the most of data the harvest already produces.

AI also brings a flexibility that mechanics cannot offer: the same architecture adapts to the type of asparagus and the type of production, whereas a mechanical system would have needed a redesign for every configuration.

Reliability and profitability: the two criteria that make or break a vision project

Two requirements framed the entire project. They apply far beyond asparagus.

• Reliability in an open environment. The machine has to harvest come rain, wind or shine, and in sand. A vision system that only works between 10 a.m. and 4 p.m. in clear weather is not a vision system: it is a demo.
• Profitability. A machine that works but that nobody can afford solves no problem. A robot only makes sense if it costs less than what it replaces over its lifetime. That is what ruled out both the overly sophisticated approaches and the overly basic ones.

Sylektis's verdict, without hesitation:

“For us, vision was the key to delivering a solution that simply works. But above all, at a competitive cost.”
Wilfried Garrigue – Co-Managing Director, Sylektis

Stuck on a detection, measurement or guidance step?

Designing a special-purpose machine where perception is your sticking point? That is exactly the kind of problem our teams work on, in food processing as well as in manufacturing.
LET'S TALK ABOUT YOUR PROJECT

What this project teaches special-purpose machine builders

Three lessons stand out, and none of them is specific to agriculture.

1. The technical bottleneck is not always where you are looking

Sylektis had mastered the mechanics, the kinematics and the integration. The sticking point was perception — a different discipline, with its own R&D cycles and its own tools. Six months spent refining feelers are six months not spent on the machine itself.
Recognising early that a bottleneck calls for outside expertise is an engineering decision in its own right.

2. Machine vision building blocks cut across sectors

This may be the most counter-intuitive point for a machine builder: technology building blocks travel from one sector to another.​​​ Defect detection on a production line, robot guidance in the field: these are different applications of the same foundation.
Expertise built on industrial quality control carries over to agricultural harvesting — and vice versa. A tough agricultural project improves the algorithms used in the factory.

3. Commitment, responsiveness, trust: the human factor

Asked what stood out in the collaboration, Sylektis named three qualities, in this order: commitment, responsiveness, trust.

On a multi-year project where you move forward with no guarantee of results, those three weigh as much as algorithm performance. It is not a polite formula: it is what keeps a project going after a technical failure.

What's next? France, Germany, Canada and other crops

The target market is France first, then Europe. By 2027, Sylektis plans to be present in France, in Germany — where large growers are already approaching them — and in Canada, whose green asparagus sector is facing a severe labour shortage.

Beyond that, the logic is one of reusable building blocks. The vision exists, the robot exists: what remains is adapting and extending them to other crops. Other vegetables. Fruit.
As for the collaboration between the two companies, it is only just beginning.

Frequently asked questions about machine vision in agricultural robotics

How does a robot detect an asparagus spear?

With 3D vision. A camera mounted at the front of the machine, above the harvesting plane, reconstructs the plot as a point cloud — a three-dimensional representation of the soil and whatever sticks out of it. An artificial intelligence algorithm interprets this point cloud to identify the spears, determine the coordinates of their tip and base, measure their height, and decide which ones are long enough to harvest. These coordinates are then sent to the cutting unit.

Can a robot-mounted vision system detect both white and green asparagus?

Yes, but not with the same sensors. White asparagus grows under the ridge and is harvested as soon as it breaks the surface: detection combines a 2D camera, which provides texture and colour, and a 3D camera, which provides relief — needed to tell a spear tip from a stone of the same colour. Green asparagus grows above ground and is cut at 20 to 30 cm: geometry is enough, and detection relies on 3D vision alone, using a point cloud.
What does not change is the software building block. The same detection foundation developed by APREX Solutions handles both cases: first designed for white asparagus, then extended to green. That is what allows the AsperCut to harvest up to 24 white spears and 40 to 80 green spears per minute, depending on the configuration. A mechanical system would have required two separate machines.

What camera does a robot need to harvest vegetables?

It depends on the task. For asparagus harvesting, the choice is a 3D camera placed above the harvesting plane: it provides the geometry of the scene, which is essential for positioning a cutting unit accurately. A 2D camera alone can recognise an object, but cannot locate it in space. The sensor, however, only does part of the job: it is the interpretation algorithm that determines how reliable the system really is.

What is the difference between machine vision in a factory and machine vision in the field?

In a factory, you detect a known object in a controlled environment: standardised parts, stable lighting, fixed distance, steady rate. In the field, every object is different — height, curvature, maturity — and the environment changes constantly: light, soil colour and texture, rain, wind, dust. An agricultural vision system must therefore be designed for variability, not repeatability.

Why isn't mechanical detection enough to harvest asparagus?

Because it is unreliable in the field and risks damaging the produce. Sylektis first developed mechanical feelers that touch the spear to signal its presence — a deliberately low-cost approach. Out in the field, it proved ineffective, and the risk of damaging the spear was a deal-breaker for growers. That is what led to the choice of vision-based detection.

Should a special-purpose machine builder develop its machine vision in-house?

Rarely, and never by default. Mechanical design and AI-based vision are two distinct disciplines, each with its own R&D cycles and learning curves. Bringing the second in-house when you excel at the first often means signing up for several years of development — while the market keeps moving. A partnership lets the machine builder keep control of its machine and its customer relationship, while integrating a perception building block that has already been proven.

Can machine vision be transferred from one sector to another?

Yes — in fact, it is one of its most useful characteristics. Technology building blocks travel: defect detection on a production line, dimensional inspection, sorting and robot guidance in the field are different applications of the same skill set. Expertise built on industrial quality control carries over to agricultural harvesting.

Who developed the vision system for the AsperCut robot?

APREX Solutions, a company specialising in machine vision and artificial intelligence applied to quality control and robot guidance. Called in by Sylektis after mechanical detection solutions had failed, it developed the 3D-camera and point-cloud detection system fitted to the AsperCut, first for white asparagus and then for green.

Stuck on a detection, measurement or guidance step?

Designing a special-purpose machine where perception is your sticking point? That is exactly the kind of problem our teams work on, in food processing as well as in manufacturing.
LET'S TALK ABOUT YOUR PROJECT

Data sources
7,395 ha in production, 27,476 t harvested in 2024   
Agreste, 2024 annual agricultural statistics — final figures, November 2025

400–500 h/ha for harvesting, ~€1.50/kg, 100–200 h/ha at the packing station, 1 FTE per 20–30 ha
Christophe Paillaugue, President of AOPn Asperges de France, 3rd technical day, Castelnaudary, reported by Réussir fruits & légumes (February 2024)

German minimum wage at €13.90/h on 1 January 2026, call for a seasonal-worker exemption
Die Weltwoche (French edition), April 2026, citing Apollo News and the Beelitz growers' association

Germany as Europe's leading producer
Converging industry sources

Covid origins, feelers, 3D camera, point cloud, mapping, 2027 markets, the three qualities of the partnership
Transcript of the APREX × Sylektis video

AsperCut specifications (areas, harvesting rates, autonomy, hybrid power)
sylektis.com, accessed 4 August 2026