Sweeper picking robot sinks its teeth into pepper harvest

Harvesting a pepper with the 'Sweeper' picking robot takes an average of 24 seconds. But a lot of innovation was needed to fully automate the picking process, says Jos Balendonck of Wageningen University & Research (WUR). At the Machine building event Balendonck will provide details about the project.

By: Dimitri Reijerman

Earlier this year the Sweeper in a commercial greenhouse. During this demonstration, the robot managed to harvest yellow peppers completely independently. That moment was a milestone in the development of the harvesting robot, a project in which a consortium with a European consortium within the European Sweepers project has collaborated for almost four years in the field of agro-robotics.

The seed for this work was planted twenty years ago at WUR, says Balendonck: “We worked on a first robot at the end of the last century, for high-wire cucumber cultivation. That was the start of the robotics adventure within our club. We have continued all that time. About eight years ago we participated in a European project, CROPS. In it we developed a first prototype of the pepper picking robot from which we learned a great deal.”

Peppers in Dutch greenhouses grow on a substrate block. The plant has three stems that grow up in a V-shape to a height of 3 to 4 meters. The Sweeper is able to harvest about half of the peppers on a plant; fruits that grow on the inside of the plant cannot yet be picked by the Sweeper. "The robot cannot yet recognize and reach them there," says the researcher. "Also, some peppers grow in clusters and leaves are in the robot's field of vision. We cannot harvest those peppers either. Getting the cutting blade in the right place, in order to cut the stem, is also a challenge."

Deep learning

The Sweeper uses a combined camera that is placed in the gripper to recognize a pepper. The camera records an RGB image of the fruit, while also creating a depth image. Using deep learning algorithms, the robot is able to recognize a fruit with great accuracy. The ambient light does not disturb the images because a very bright flash is used. Using deep learning algorithms, the robot is able to recognize not only a fruit, but also the stem, leaves and stalks. These deep learning algorithms have been extensively trained and tested.

A vibrating blade has been developed for cutting the fruit. positioning it so that the stem of the pepper cuts as close as possible to the plant stem without damaging the plant and pepper, Balendonck explains.

While the first prototypes of the picking robot needed more than a minute to harvest a single pepper, the current Sweeper is able to do that in 24 seconds. But a human picker only needs 5 seconds. The WUR researcher thinks that the speed of the Sweeper can be doubled by parallelizing the two main tasks – picking and moving. Other parts of the automated picking process can also be optimized.

Market introduction

Balendonck thinks he can launch a market-ready robot in about three to five years: “The market introduction will take place in two steps. First, the robot will work together with humans, with each picking about half. If we adapt the cultivation system to a single-row setup and the picking robot can get behind the plant in this way, the robot can become fully autonomous, possibly in five to ten years.”

Jos Balendonck will be present on December 6 during the Machine building event An give a keynote about the Sweeper. You can watch this event after registration free visits.