Why orchard robots need a job beyond the harvest

For Gilbert Plath, the arithmetic of growing fruit has changed. The technology director at Washington Fruit and Produce told Cornell that labour accounted for about 45% of his company’s production costs 15 years ago; now it exceeds 60%. That is one grower’s account, but it explains the appeal of a machine that can take on more orchard work.[1]

Cornell announced a four-year, US$7.5 million, USDA-funded orchard-robotics centre on 3 September. Its ambitions span pruning, thinning, pollination and harvest. The commercial idea is compelling: earn a return on the same expensive platform through several seasons.[1]

The hours that pay for the machine

Owning machinery brings costs even while it sits idle. Iowa State’s machinery-cost guide separates annual ownership expenses, including depreciation, financing and insurance, from operating expenses such as fuel, repairs and labour. For an orchard robot, the productive-hours denominator can matter as much as the price on the invoice.[2]

Consider a deliberately hypothetical machine with US$60,000 in annual ownership costs. Spread across 300 productive hours, that is US$200 an hour; across 900 hours, it is about US$66.67. Those are calculated allocations of the same fixed cost, not quoted robot prices or a forecast of profit. Extra attachments, supervision, maintenance and wear still have to be paid for.

This gives AI developers a different target from winning a picking-speed contest. A platform needs transferable capabilities: seeing through a cluttered canopy, approaching a target, and controlling a tool precisely enough to perform useful work. Each additional task must bring in more value than the equipment and effort required to add it.

A pruning tool has to learn the tree

A separate September preprint by Abhinav Jain, Cindy Grimm and Stefan Lee illustrates the engineering challenge. Their system learns in simulation to steer a pruning tool using visual feedback, then transfers the policy to physical trees. Its training combines automatically generated example movements with reinforcement learning, in which simulated attempts help refine the strategy.[3]

In outdoor trials, the tool reached the required cutting position in 10 of 14 apple-tree attempts and five of 14 cherry-tree attempts. The cherry planting differed more from the geometry used in training. These were selected, reachable targets; people positioned the mobile base and could intervene. Success meant aligning the branch in the cutting mechanism near the designated point, rather than autonomously deciding how to prune an entire tree. Successful field and laboratory approaches averaged about 41 seconds.[3]

The difference between the two orchards is revealing. Sharing a camera and robot arm does not make every growing system interchangeable. A business built on changing tools between seasons must also account for changes in tree shape, task geometry and the skill the software has learned.

The value of an unbruised apple

More working hours only help if the work is valuable. A 2025 economic study led by Diane Charlton used a 2019 Washington Gala-apple budget to model robotic harvesting. Its baseline assumed nine arms, five seconds per apple per arm, robots picking 60% of the fruit and people collecting the rest. With one operator supervising two robots, 40.5 acres harvested annually per robot, and ten-year financing at 5%, the affordable purchase cost was about US$79,046: the price at which the model matched the profit from hand harvesting.[4]

One change transformed that figure. Reducing the robot’s additional fruit-damage rate relative to hand picking from five percentage points to one, while holding those other assumptions fixed, raised the affordable price to US$224,317. These are modelled thresholds, not vendor quotations or 2026 orchard forecasts. The mechanism is straightforward: more fruit retains its fresh-market value instead of being downgraded for processing. Improvements in handling can therefore justify substantial investment, even without picking another apple per minute.[4]

That makes seasonal versatility a demanding proposition. The pruning attachment must produce a worthwhile cut, the thinning system must remove the right fruit, and the harvester must preserve what the grower can sell. A busy robot can still be an expensive robot if its work creates losses elsewhere.

Fresh Fruit Robotics founder Avi Kahani told Cornell that growers had questioned the return on a costly machine used only for harvesting. Their feedback favoured equipment that could also prune, pollinate and thin. The new centre includes economists and sociologists alongside engineers and horticultural researchers—an acknowledgement that adoption involves more than making an arm move.[1]

The practical question for a grower is what an additional season of robotic work earns after its own costs. Answering that will require a calendar of useful tasks, crop-quality results and accounts from operating machines. The most persuasive orchard robot may be the one with a full order book for the year.

Report an error or suggest a correction