Agri-OpenCore finishes with a tomato demo and reusable robot software

Agri-OpenCore ended its project meetings with a robot harvesting tomatoes on the vine at APS Produce’s Europa Nurseries in Kent. The demonstration took place on 19 October, according to the consortium’s account published two days later. Alongside that visible result, the project leaves a less conspicuous collection of tools: open software that other teams can use to build agricultural robots.[1]

The programme ran from April 2023 to September 2026, bringing together APS Group, the University of Lincoln, Dogtooth Technologies and Extend Robotics with Innovate UK funding through Defra’s Farming Innovation Programme. It worked on strawberry and high-wire tomato harvesting. Project leader Phil Pearson described strawberry robots operating commercially while identifying tomato harvesting as work still progressing towards commercialisation.[1]

The reusable part becomes clearer in the released architecture. Agri-OpenCore separates a robot’s hardware configuration, its farm environment and the algorithms doing jobs such as perception or navigation. A complete scenario brings those pieces together. That separation gives a developer somewhere specific to make a change: a new sensor need not require redesigning how the farm’s routes are represented.[2]

What a fruit detector actually supplies

The public AOC Fruit Detector is one concrete component. It uses Mask R-CNN, a computer-vision method that separates individual objects within an image, to identify strawberries or tomatoes. With colour and depth information from a camera, the package can report fruit location, size and ripeness. Pretrained models are provided for both crops, alongside a procedure for refining them or training for another crop.[3]

The code documentation exposes assumptions that a glossy demonstration can hide. Accurate position estimates require camera calibration, suitable colour images and depth aligned with them. The package also permits a fixed assumed distance when depth is unavailable, defaulting to one metre. That fallback is a configuration choice, not a measurement of how far away each fruit really is.[4]

There is also a route for examining predictions before connecting the package to a live robot. Its documented non-ROS mode reads images from a directory and saves annotated pictures and JSON records. In ROS mode, it receives camera data and publishes fruit information and visualisation outputs. The same perception component can therefore be inspected as a producer of data before it becomes part of a machine’s actions.[4]

The tomato demonstration and the public detector are separate documented outputs of the project. The close-out account does not identify every component used in the demonstrated machine, so the released package should not be treated as a complete reconstruction of that robot.[1]

A farm that can change before the robot arrives

The tomato simulation tools supply another reusable layer. They generate glasshouses and plants in Gazebo and Unity, with adjustable row counts, spacing, lighting and farm size. Individual plants vary in stems, leaves and fruit. A team can therefore change the layout or crop appearance in a simulated setting while developing the software that will encounter those differences.[5]

Navigation has a separate field record. The project reports that a Dogtooth robot travelled autonomously for 13.5 kilometres over seven operating days at Annabel’s Farm between June 2025 and August 2026. Those trials took place in commercial strawberry polytunnels. The distance measures navigation, not tomato picking speed. An accompanying open logger records autonomous distance and interventions, including distance between interventions.[6]

Agri-OpenCore says its core components are open source, while selected components were retained for commercial exploitation. That makes the deliverable an open foundation companies can extend, rather than a promise that every part of a commercial harvester is available. The architecture also packages development environments into containers, helping teams keep their software setup consistent as components move between development, testing and deployment.[6][2]

For another robotics team, the practical gain is a more explicit starting point. It can inspect the fruit data, substitute a hardware configuration, vary the simulated farm and see where its own crop or machine needs additional work. The tomato demonstration gives the project an ending; the documented interfaces give its components a way to continue in someone else’s system.

Report an error or suggest a correction