Harvesting robots and AI sorting in mushroom and soft fruit growing: what works and what does not in 2026

Three things that are often lumped together

In discussions about robotisation, three different technologies are often confused, even though they differ greatly in maturity and cost. If you keep them apart, you are much better placed to judge what is feasible for your own business.

It pays to ask of each technology: where does this stand in practice, and what does it really cost in my situation?

• Vision-based ripeness scanning: a camera plus an AI model that recognises shape, colour and growth rate and estimates when a fruit or mushroom is ready to pick. This is mainly measurement and prediction, not a moving robot.
• Quality sorting: a vision module that recognises defects after harvest, such as overlong stems, open mushrooms, damage or soiled stem bases, and sorts them out or returns them to the line.
• Picking robots: machinery that actually locates, assesses and harvests the fruit or mushroom by itself, without a human hand. This is technically the hardest and the furthest from widespread use.

Mushrooms: sorting works sooner than picking

With mushrooms, the most tangible gain at the moment lies not in the picking robot but in AI-assisted quality sorting. Dutch projects are investigating whether vision and AI can automatically assess machine-harvested mushrooms for defects: soiled stem bases, overlong stems, open specimens or damage. Today that assessment is done by hand, which takes a great deal of labour and is prone to differences between employees. A vision module that pre-sorts consistently takes over part of precisely that repetitive, labour-intensive work.

The picking robot for mushrooms is a different story. Work on it is under way internationally, including by Canadian and European organisations, but picking mushrooms without bruising them, in a dense bed and in the right order, is technically difficult. Treat announcements about it as a direction of travel, not as something that will be in your growing room next season. In most businesses, the shortest route to results runs through sorting and quality control, not through fully replacing hand picking.

Soft fruit: promising, but it remains largely human work

Strawberries and other soft fruit are small, soft and variable within the same batch, and that makes fully automated harvesting difficult. Strawberry harvesting robots have been in development and trials for years, and research by Wageningen University & Research, among others, focuses on determining harvest ripeness for each individual fruit, precisely because the variation within a batch is large with soft fruit. The technology is moving forward, but the sober conclusion from practice is that robots do cut and pick, yet it remains largely human work.

For soft fruit growers, the most honest question in 2026 is therefore not whether a harvesting robot replaces the picking crew, because it does not yet do that, but where technology supports people. Ripeness prediction that sharpens harvest planning and staffing, or quality sorting after harvest, often delivers returns sooner than a fully automated picking robot that still has to prove, under your specific conditions, that it can cope with the pace and the softness of the fruit.

Honest about costs and payback periods

A harvesting or sorting robot is a substantial investment, and the payback period depends on factors that vary greatly between businesses: how many hours of manual work it really replaces, how many days a year it runs, what maintenance and breakdowns cost, and how much supervision and corrective work is still needed. Field research into robot business cases mainly shows that this calculation is specific to each business and that a generic payback percentage says little about your situation.

So be critical of savings claims. A vendor who quotes a fixed payback period or a round savings percentage without knowing your own figures is selling an assumption. Ask for a trial or pilot with your own product, your own variety and your own line speed, and work through the business case with your actual labour costs and operating hours. Preferably start with the cheapest and most proven step, usually vision sorting, before you look at a full picking robot.

Bear in mind, too, that supervision is often underestimated: the people who configure, check and adjust the system must know what it can and cannot do. A focused training day belongs in the business case. Our Practice Day is one such day, in-company on your own process, from €2,500 per group.

The human decides and signs

A vision model gives a judgement based on what it has learned to recognise. It does not know your agreements with the buyer, does not know that a particular batch is behaving differently today, and only sees a new defect once it has been trained on it. That is why the grower or quality manager remains the one who decides and signs, even when a camera pre-sorts. For businesses that work under quality and certification schemes, AI changes nothing about final responsibility for quality and record-keeping.

In addition, the AI literacy obligation in Article 4 of the EU AI Act (Regulation (EU) 2024/1689) has applied since 2 February 2025. Since 27 July 2026, that provision has asked for measures that support the AI literacy of the employees involved. In practice, anyone deploying an AI sorting system or harvesting robot makes sure those employees understand how the system arrives at its judgement, where the assumptions and limits lie, and when an outcome must be adjusted manually. That is not only a legal matter but also a practical one: a system that nobody understands is either followed blindly or ignored, and both are expensive.

Where to start in practice

Do not start with the question of which robot to buy, but with the question of which process currently costs you the most labour, money or hassle. Often that is manual quality assessment or estimating the right moment to harvest, and that is exactly where the step to vision support is smaller and cheaper than the step to a fully automated picking robot.

First record how things work now: how many hours of manual work, which defects are assessed and by whom, and what the decision to harvest is based on. That baseline measurement lets you judge a pilot or vendor trial fairly on what it really delivers, instead of on a promise. Only when one step proves itself in your own figures does it make sense to expand.