Test: weeding robots capable of removing majority of weeds

To find out how effective two early stage development weeding robot prototypes are in detecting and removing weeds in practice, the Dutch Farm of the Future conducted initial tests on request of the Dutch National Fieldlab for Precision Farming (NPPL). The Farm of the Future is part of Wageningen University & Research (WUR) and the tests were done by their researcher Kees Meesters.

The invitation to take part was welcomed by the Swedish Ekobot team with their WEAI autonomous weeder project and the Dutch Odd.Bot team with their Quirky weeding robot (see boxes for details and manufacturers’ feedback to the test results).

Setup and test procedures

Because both participants are initially focussing on different crops, they performed the tests in the type of crop of their choice: Ekobot in onions sown/planted on beds and Odd.Bot in carrots sown in ridges. Both manufacturers were allowed to trial their robot is the designated crop and to (re) set their technology in the best way possible prior to the tests.

The initial idea of executing the tests with planted standardised artificial weeds made of plastic, had to be abandoned because the Odd.Bot Quirky is (currently) specialised and trained to detect specific types of weeds in carrot crops. The plastic weeds couldn’t be detected by its camera(s).

_{Text continues below image}

Marked plots

The test areas consisted of marked plots of each field in which researcher Kees Meesters counted and noted down every natural weed present (each plant that was not the cultivated crop). These weeds varied in size (width and height) and in location. Some of them were close to the crop, others were further apart.

The Ekobot WEAI attempted to remove the weeds on the onion bed between the rows and in the rows to within (currently) 2 cm from the onions. The Odd.Bot Quirky attempted to remove all the weeds on top of the horizontal part of the ridges. Each test was performed three times on three separate plots.

The effectiveness of the weeding jobs was determined by checking and counting the number of weeds removed by the robotic arms of the robots. ‘Removed’, in this case means that there had to be no physical signs left above ground of the weeds present prior to doing the tests. These evaluations were done visually. No digging was done and no checks on the results were performed in the days and weeks after the tests.

The driving speeds and the capacities of the two tested robots were not relevant for the purpose of the tests and therefore not determined nor measured.

31 to 81% of weeds removed

The results of the tests are expressed as percentages of the numbers of weeds removed compared to the numbers of weeds present prior to the tests. Both the Ekobot WEAI and the Odd.Bot Quirky each showed their performance in three different plots (so six test plots/areas in total).

The absolute number of weed plants in each of the six plots prior to the test is irrelevant as it’s the percentage of weeds removed that counts. Because of the non-standardised test procedures and because robots worked in different crops, it is not possible to compare the test results of both robots.

_{Text continues below image}

Individual results vary quite a lot

The table below shows the individual results as well as the average scores per robot and of the whole test. The individual results vary quite a lot: for the Ekobot WEAI between 30.8% and 64.5% of weeds were removed and for the Odd.Bot Quirky between 33.3% and 81.4%. On average, Ekobot scores 47.1% and Odd.Bot 64.2%.

“Statistically there’s no significant difference between the results of both robots. Neither of them damaged the crop so that’s positive as well. I did notice that both robots were struggling to detect a specific (unknown to us) red type of weed but I can’t really explain why”, says Kees Meesters.

If a general conclusion is to be drawn, then one could say that the robots manage to detect and remove 55.7% of the weeds present in the test plots in one run.

Looking back, Kees Meesters reflects on the tests and the circumstances as follows: “We first of all learned that the initial idea of using artificial weeds – a quite common approach in research to standardise circumstances – doesn’t work because of the sophistication of the algorithms. While testing, I had the idea that the size of the weeds affected the number of weeds detected by the robots.”

“For future tests, I would like to have the ability to control the type, the number and the size of weeds. Maybe we can sow them at different moments in time. A second wish, both of myself and of the participating robot manufacturers, is to perform the tests on various types of soil. In this case, we did the tests on a light clay type of soil.”