Polybot teaches a robot to harvest tomatoes by watching humans work

A greenhouse is not a perfectly repeatable assembly line. Two clusters can have different shapes; leaves hide the cut point; light shifts; the fruit moves when the tool touches it. Resolving every exception with a new rule consumes engineering time and can produce a fragile system.

Polybot is testing another approach. An operator carries out the harvest, and the system records the relationship between what the robot sees and the action performed. According to co-founder Martin Kiefel, the machine receives demonstrations of the movement, but no one needs to explicitly describe what a tomato is. With enough examples, the model learns regularities: the red color, the geometry of the cluster, and the right moment to cut.

This does not mean the robot understands the tomato like a human does. It means it learns an action policy from visual and motor data, rather than relying solely on hand-written conditions. The difference is significant: teaching new examples can be faster than reprogramming an extensive collection of individual cases.

The timeline reported by Polybot illustrates the speed of development. Laboratory work began in January 2025. In May, the machine autonomously picked its first tomato in a real greenhouse. By December, the company claimed it could reliably navigate and harvest full rows.

The next planned step is a fully integrated pilot with growers during the summer of 2026. The company aims to deliver the first commercial systems in early 2027. These last two dates are corporate targets, not completed milestones, and will depend on sustained performance outside of a controlled demonstration.

Polybot emerged from the environment of the ELLIS Institute Tübingen and the Max Planck Institute for Intelligent Systems, with support from the Tübingen AI Center and the Cyber Valley network. The initial idea is linked to researcher Wieland Brendel; the founding team includes Martin Kiefel, Claudio Michaelis, Maike Kaufman, and Sebastian Blaes. Georg Martius and Brendel serve as advisors.

In 2025, SPRIND—the German Federal Agency for Disruptive Innovation—awarded the project a seven-month validation contract for approximately 220,000 euros to facilitate the transition from research to business. Upon incorporating as a spin-off in December, the project announced an additional one million euros in support from SPRIND.

Public funding allows for the testing of a technology that combines computer vision, imitation learning, a robotic arm, and a platform that runs on the cart rails already installed in many greenhouses. Leveraging this infrastructure removes a practical barrier: growers would not have to completely overhaul their operations to introduce the machine.

Harvesting is not just about locating something red. The machine must approach without knocking into other fruits, identify the correct stem, position the tool, cut, and deposit the cluster without damaging it. It must also distinguish ripeness, operate amid occlusions, and repeat the cycle for hours with an economically acceptable error rate.

Polybot claims to have reached an approximately human speed and aims to increase it by a factor of 1.5. It also envisions related tasks such as plant de-leafing, winding, and tying. There are still not enough public metrics on fruits per hour, percentage of correct cuts, damage caused, human interventions, energy consumption, or availability over a full workday.

The absence of these figures does not invalidate the visible progress, but it does limit what can be concluded. The video shows a concrete capability; a commercial operation will have to demonstrate repeatability, safety, easy maintenance, and an overall cost that is lower than or competitive with manual labor.

Polybot's interest goes beyond tomatoes as a crop. Many tasks were never automated because their volume did not justify months of programming or because the environment changed too much. If an experienced worker can generate data by teaching the process, the cost of adapting a robot to local variations could drop substantially.

The same logic is relevant to short-run manufacturing, logistics sorting, visual inspection, order picking, and warehouse operations. However, these extensions are an interpretation of the potential of learning by demonstration: Polybot has only presented public evidence in agriculture and has not announced commercial deployments in those other sectors.

The company has built a compelling narrative and a visible demonstration: learning from humans, rapid progression, and the use of existing agricultural infrastructure. To turn it into a product, it will need to publish or allow verification of data on productivity, damage, error recovery, maintenance, safety, return on investment, and performance across varieties and seasons.

If it passes that test, its greatest contribution may not be a tomato harvester. It could be a method for automation to reach complex tasks where writing out every step was always too slow and costly. In that scenario, the most useful language for programming an industrial robot would simply be showing it how the work is done.