Tom Scott’s new video is a visit to Tiptree, the Essex farm behind the jam, to find out why robots still can’t pick strawberries. His conclusion, updated a year after filming: he could find tech demos and trials, but no machine picking strawberries faster and cheaper than people at commercial scale.
The video is a list of things at Tiptree that do work. Watching it, I noticed they all work the same way, and the video never says so.
The strawberries are in the air. The third-generation system lifts plants off the ground into troughs of crushed coconut husk. The farm manager, Andrey, says what that buys: “We stop applying herbicides.” Nothing grows in coconut husk except what you plant in it. The weeds weren’t beaten. The plants were moved somewhere weeds can’t follow.
The fungicide is a robot, and it works. An Antobot machine drives the rows at night shining UVC on the plants, three times a week, with cameras counting buds and berries for a yield report by morning. Roughly seven of them are running. They work because the rows are regular, and the farm intends to keep it that way. In Andrey’s words: “With the new systems, with the new greenhouses we are building, we’re making them much more easy for the robots to operate.”
The heaviest machine does almost no work. In the fourth-generation greenhouse, alternate rows of plants, 40 metres long and “about ten tons per bay”, swap height at the press of a button so pickers can reach both. One motor drives the whole set, because the rows are counterweighted against each other. “So, basically, the motors are doing very little work. They’re only rotating.” The physics of the structure does the lifting.
The picking robot was designed the same way. The University of Essex team’s prototype has no camera or sensors on its gripper. A shaped stem holder steadies the stalk and a soft pad catches the cut berry. The researcher put it in one sentence, which is the title of this piece: “We want to have the structure perform the function.” Tom liked the line. So do I. It describes every success in the video.
So why is picking the one job still undone? The video’s explanation is that it’s hard, which is true but doesn’t say which part is hard. The published numbers do.
A 2025 review of deep-learning picking robots (Sensors 25(12), open access) tabulates results from many papers. For strawberries, the detection and localisation entries are mostly very high: 89% to 99%, with one model at 99.8%. Finding a ripe strawberry in a picture is close to a solved problem.
The one strawberry entry in that table that reports an end-to-end picking success rate says 65.1%. It’s from Xiong, Ge and From’s 2020 paper on picking fruit in clusters, whose whole subject is the moment the video shows: a target berry surrounded by others, leaves and stems. Their robot pushes the neighbours aside, in a zig-zag, or drags the target somewhere clearer and pushes back, before it can cut. Tom watched the Essex robot do something very like this and said what startled him was that it “managed to move all of the other ones out the way.”
So the gap isn’t in the eye, it’s in the reach: a soft object in a tangle, with other soft objects in the way, that bruises if you touch it wrong. That’s exactly where “make the robot cleverer” is most expensive. The video’s own number is a million-pound robot with fingertip touch, and touch “still not to the level that you can get precise feedback from it”.
The tomato industry ran this experiment once already. In California in the 1940s and 50s, an engineering team at UC Davis built a mechanical tomato harvester, and it crushed the tomatoes. The fix came from the plant breeders. Jack Hanna bred a firmer, blockier variety, VF145, that could survive the machine and wouldn’t roll off its conveyor belts. Within a few years nearly the whole California tomato industry was using mechanical harvesters. They didn’t build a harvester gentle enough for the tomato. They bred a tomato tough enough for the harvester.
Tiptree is already doing the strawberry version of this, one layer at a time. The plants are lifted to a height a machine can drive under, in rows a LiDAR can follow, in greenhouses being designed for robots. Breeding is on the list too: Tom mentions that some strawberry cultivars are “optimised for hardiness, they don’t bruise during transport”. Every one of those changes moves the problem from the machine into the structure around it.
My guess, and it’s only that: the first strawberry picker that works commercially won’t be the one with the best hand. It will be the one working in a growing system shaped so that each berry hangs alone, at a known height, with a clear stem. If that happens, the Essex team’s sentence was the right one, it just needed to be applied to the whole farm instead of the gripper.
There’s a cost I don’t want to skip. The berry Tom ate in the fourth-generation greenhouse, an Aurora Karima straight off the plant, was “the best strawberry I’ve ever tasted”. A berry bred to hang alone and survive a machine is a different breeding target from that one. Whether those two can be the same berry is the question I’d most like a breeder to answer.
Notes. Quotations from the video are checked against its English transcript. The 89–99.8% range and the 65.1% success rate come from the results table in the Sensors review (PMC12197199), read in full via Europe PMC. I read only the abstract of Xiong, Ge and From 2020, so the 65.1% is as the review reports it, and the paper’s own test conditions may differ from Tiptree’s. I tried to read the same group’s larger 2020 field evaluation in the Journal of Field Robotics, which is open access, but the publisher, a browser fetch and both repository copies refused my requests, so it isn’t used here. The tomato history is from Wikipedia’s article on Hanna, which is enough for the outline and inconsistent on exact years, so I’ve kept only the decades.