Picking Fruit Is Not the Same as Harvesting It

A robot lifts an apple from a branch. The movement looks successful. But did the fruit arrive intact, how much of the crop could the machine reach, and what happened when the apple was inspected later?

Agricultural robotics needs a broader definition of success than a clean-looking grasp. The useful outcome is fruit collected in saleable condition, not simply fruit moved. Three research examples show why those measures should stay separate.

Keep the denominator attached

A June 2026 preprint reports trials of a dual-arm apple harvester in two commercial orchards during the 2025 season. Across 1,738 arm cycles, the authors report 80.0% success per attempt and a mean per-arm cycle time of 7.53 seconds. These are attempt-level and arm-level measures, not a statement that the machine harvested 80% of every orchard's crop.¹

The paper's separate Michigan quality assessment found 271 of 297 robotically harvested apples, or 91.2%, remained Extra Fancy. That grading sample is not the 1,738-cycle trial denominator. The result also varied by cultivar: 98.0% for EverCrisp and 84.7% for Pink Lady. Platform positioning still relied on a human operator.¹

This is useful progress precisely because the report exposes more than one result. Combining unlike percentages into a single headline would hide the conditions that make the system useful. A sensible comparison asks what was attempted, what was collected and what passed the relevant quality check.

Inspect after the impressive moment

Damage assessment needs a clock. A tomato that looks unmarked immediately after handling may not produce the same result at a later inspection.

The October 2025 ANGEL preprint describes a cable-driven soft gripper that closes compliant pockets around fruit. For tomatoes within its effective size range, the authors report zero immediate damage and less than 9% bruising after five days.² The two observations are not interchangeable. The delayed result is a different measurement, not a contradiction to erase from the story.

The practical implication is to specify both the inspection method and its timing. An immediate visual check, a later bruise assessment and a commercial grading decision answer different questions. These author-reported studies do not establish a universal damage rate across crops or an independently replicated performance guarantee.

Separate the gripper from the whole robot

A 2023 blackberry study makes another important distinction. Its field tests harvested 139 berries with manual positioning. Separately, integration with a UR5 arm demonstrated visually guided picking of fifteen artificial blackberries in a laboratory.³

Those tests contribute different evidence: handling real fruit in the field, and moving the integrated system toward targets in a controlled setting. Neither result should be relabeled as autonomous field harvesting of 139 berries. This is not a reason to dismiss component research; it is a reason to describe what each experiment actually establishes.

A better harvesting scorecard

For a practical evaluation, report the chain rather than one success number. Start with the crop available to the system, record fruit detected and reachable, then count attempts, completed collection and acceptable fruit after a specified inspection interval. Include missed fruit, retries and human interventions instead of making them disappear from the denominator.

Measure time over the same boundary. A single arm's cycle time should not be treated as the throughput of an entire orchard operation. Repositioning, collection handling and interruptions belong in a whole-system trial even when a component benchmark deliberately excludes them.

This scorecard is an editorial recommendation, not a new experiment. Its purpose is straightforward: make it possible to see where improvement is needed. A faster grasp helps only if the rest of the process preserves its value. The strongest harvesting story ends with the condition of the crop, not the motion of the arm.

References

1. Zhu and colleagues. A Modular Dual-Arm Apple Harvesting Robot with Enhanced Field Performance. arXiv preprint, version 1, June 12, 2026. Field trials concern the 2025 season; immediate grading and delayed inspection use separate samples. Sections 4 and 5 reviewed October 18, 2026.
2. Patel and colleagues. ANGEL: A Novel Gripper for Versatile and Light-touch Fruit Harvesting. arXiv preprint, version 1, October 20, 2025. Abstract-level results for tomatoes within the effective size range; not an independent replication. Accessed October 18, 2026.
3. Qiu and colleagues. Tendon-Driven Soft Robotic Gripper with Integrated Ripeness Sensing for Blackberry Harvesting. arXiv version submitted February 6, 2023; the authors identify the paper as accepted to ICRA 2023. Field/manual-positioning and laboratory/artificial-fruit tests are distinct. Accessed October 18, 2026.