Agree the trial before the demonstration
Start with the grower's decision. A research team may want to establish that a new cutter works. A buyer needs to know whether the integrated system can deliver useful produce within a harvest window. Both are legitimate questions, but they require different evidence. State the objective, crop specification and operating conditions before selecting a machine or interpreting a result.
Choose representative rows and document any preparation. Leaf removal, simplified crop training, hand-selected targets and restricted weather can improve a demonstration without proving broader readiness. Rather than rejecting those conditions automatically, record them as dependencies. The adoption question then becomes whether the farm can maintain them economically and consistently.
Keep the denominator visible
“Success rate” can mean success among attempted picks, among detected targets or among all eligible crops in the test area. Those definitions can produce very different impressions. Ask for the count at each stage: eligible crop, detected crop, proposed target, attempted pick, detached crop and accepted produce. Record why targets leave the chain.
Eligibility should be checked independently enough to expose skipped produce. If the robot defines its own target set and is scored only on that set, difficult crops may disappear from the result. A grower-approved audit can distinguish a sensible decision to defer an unreachable crop from a perception error or a systematic exclusion of valuable produce.
Use more than one clock
A fast arm movement is not a complete harvesting cycle. The system may spend time searching, adjusting position, checking maturity, recovering from an error, placing fruit or waiting for a container. Report the boundaries of any cycle-time figure, and keep total elapsed trial time alongside it. Do not silently remove failed attempts from the average.
Availability also needs a clear definition. A machine can be powered on yet unable to harvest because it is waiting for assistance or operating outside its permitted conditions. Separate planned service, charging, fault recovery and external delays. This makes it possible to see whether a bottleneck belongs to the machine, the crop or the support operation.
Measure the human work that remains
Autonomy changes work rather than automatically removing it. People may supervise, replace containers, clean sensors, clear crop tangles or approve a restart. Log the reason, duration and skill requirement of each intervention. An operator who must watch continuously is a different labour proposition from one who can safely perform other tasks.
Do not extrapolate multi-machine supervision from a single quiet demonstration. Concurrent faults, travel between machines and access restrictions can change the workload. Remote assistance also depends on communications and clear authority. Define who can stop the system, who may enter the working area and who is responsible for returning it to service.
Failure mode: the trial ends before the hard conditions arrive
A short run may avoid full containers, changing light, heating, contamination and operator fatigue. Extend evaluation across the conditions that matter to the proposed harvest window, and report the limits of what was observed rather than projecting uninterrupted output.
Build the economic case from local inputs
Useful cost comparisons include installation, crop preparation, training, supervision, energy, maintenance, consumables, service access and finance. They should also include produce left behind or downgraded. Purchase price divided by a best-case picking rate is not a reliable cost per marketable unit.
Use scenarios rather than a single payback promise. Vary the available harvest period, eligible crop density, support labour and downtime assumptions. Consider what happens if the machine is unavailable during a peak picking window. A credible supplier should help explain these dependencies without presenting a laboratory result as a guarantee for every farm.
Finish with a bounded decision
A useful outcome may be a limited deployment on one crop layout, another trial after a tool revision or a decision to wait. Record the conditions under which the evidence supports the choice. Keep unresolved risks visible and assign an owner to each follow-up action. The objective is not to prove that all harvesting can be automated; it is to make a defensible decision about this operation.
Sources and further reading
- Bac et al., 2017, “Performance Evaluation of a Harvesting Robot for Sweet Pepper.” A crop-specific example of structured system evaluation, not a cross-crop benchmark.
- Bac et al., 2014, state-of-the-art review and challenges. Context for the dependency between harvesting subsystems.
- ISO 12100 concerns machinery risk assessment and risk reduction. ISO 25119 concerns safety-related control systems for agricultural machinery. Applicability requires qualified review.
- Natural Resources Institute, University of Greenwich, and Rothamsted Research offer wider crop and agricultural research context.