Ecorobotix SA designed and produces the ultra-precise smart spot sprayer ARA (Tanner et al., 2021; Tanner, 2022; Ecorobotix, 2023). ARA distinguishes itself in the market by combining multiple advantages. It can apply the product with great precision. The ARA technology can be transferred to any crop once a training data set is available. ARA is both ambivalent (targeting weeds or crop plants) and polyvalent (offering various weeding strategies across a range of crops, nine so far). ARA performs consistently in both transplanted and sown crops, being flexible and independent of the crop's growth stage, with the restrictions of a 2-cm minimum diameter and a 50-cm maximum height of the crop. ARA comprises three main parts (Figures 1A, B): a front tank for the mixture, an onboard tablet for machine control (Figure 1C), and the 6.04-m-wide sprayer itself. The smart sprayer includes (1) 12 Light Emitting Diode (LED) spotlights for uniform field lighting, 6 high-resolution RGB cameras, and 6 in-depth cameras, together constituting the information acquisition part; (2) 6 processors, constituting the AI core for real-time analysis of the live images and direct decision-making; and (3) a scalable boom with 156 nozzles spaced 4 cm apart, each controlled by solenoid valves for the high-precision actuation spray.

The ARA technology relies on supervised machine learning with algorithms trained on a large database of crop pictures acquired in real conditions in the field. To ensure robust decision-making, the onboarded algorithms on ARA exhibit F1 scores for weeds >90%.

During operation, ARA captures live images of the field, which are processed by its AI to classify each pixel as crop, soil, or weed (Figure 1D). The required mixture volume is then calculated and synchronized with the nozzle control, reducing the spray area to a 6 cm × 6 cm minimum on the selected target (crop, soil, or weeds). The product volume can be adjusted within standard sprayer ranges and adapted to the tractor's speed for efficient weeding. ARA's design allows it to operate close to the crop, reducing drift by more than 95% (DGAL/SDSPV/2023-282, 2023; van Steenbergen et al., 2024). This feature is essential as it minimizes phytotoxicity, protects the environment, and promotes biodiversity.

Conventional broadcast spraying solutions typically spread pesticide diluted in 150–300 L/ha of water (note that water-diluted herbicides will be designated, hereafter, as “herbicide volumes”).

However, ARA's spot-spraying methodology, which applies products only where needed, significantly reduces these inputs. Data from our ARA dashboard, extracted and anonymized, show that among 57 ARA machines performing 1,682 weeding missions on onion fields over 1 ha in 2023, 78.9% of the herbicide volume was saved across 8,266 ha (8.3 kha) compared to traditional broadcast treatments. Remarkably, one third of the treated area saved more than 90% of the herbicide volume (Figure 2A). Similarly, in sugar beet fields, 79.8% of the herbicide volume was saved (33 ARA machines, 480 weeding missions over 2 kha), with 47.6% of the treated area saving more than 90% of the volume (Figure 2A). Those two real-world examples advocate ARA's ability to drastically reduce herbicide inputs in the field, providing significant savings and environmental benefits for end users.

Both mechanical and traditional chemical broadcast weeding process the entire field uniformly, applying maximum efficacy whether weeds are present or not. In contrast, spot-spraying weeding methods, like ARA, target only specific areas where weeds are present. Weeding efficiency can be compromised if the spot sprayer algorithm is not accurate or the applied mixture is not suitable for the specific weed species. To evaluate the performance of ARA's algorithm, plant coverage and treatment efficacy were evaluated in the tested modalities over time and compared to the non-treated control modality and the tested conditions. Briefly, the ARA selective modality corresponds to a concentrated stringent chemical mixture, ARA's non-selective modality explores the efficacy of pelargonic acid at 16 L/ha, and the broadcast farmer reference modality corresponds to conventional mixtures and dosages. Information on the programs applied in different modalities is available in Tables 1–3.

Despite the low quantity of chemicals applied in ARA modalities in the onion trial (Figures 2B, C), the ARA selective modality displays a global efficacy of 92%, ARA non-selective 87% (compared to 99% for the broadcast treatment; Figure 3A, Supplementary Figure S1A). These results are satisfying (Figures 3A, B, Supplementary Figure S1A). Although ARA modalities (both selective and non-selective) showed slightly lower efficacy than the broadcast treatment, no significant differences in weeding efficacy were found.

Notably, ARA required fewer interventions to achieve comparable weeding efficacy than the broadcast program (5 vs. 9 treatments; Table 1) and used lower volumes and fewer molecules amounts than the broadcast treatment (Figures 2B, C, 3B, Supplementary Figure S3A). This suggests that ARA's onion algorithm performs well, and the mixtures applied were strong enough to control the weeds present in this trial.

Examining individual weeds, the ARA non-selective modality shows some difficulty in controlling Polygonum persicaria (POLPE, 86%). In comparison, the ARA selective modality controlled 95.75% of the POLPE and 100% for the broadcast modality (Supplementary Figure S1A). This confirms a previous study showing that pelargonic acid has difficulty controlling POLPE (Loddo et al., 2023). The control of Senecio vulgaris (SENVU) reached only 66% in the ARA non-selective modality (compared to 97.5% in the ARA selective and 100% in broadcast modalities; Supplementary Figure S1A). This can be explained by the high density of SENVU (40% in the NT) and indicates that TD should have been applied earlier if conditions had been favorable and that an extra treatment may have been necessary to compensate for the pelargonic acid's low efficacy on older weed stages. Overall, these results are very satisfying, demonstrating that ultra-precise smart spot spraying can be an effective alternative to broadcast treatment in controlling weeds in onion crops with both selective and non-selective programs. As expected, the weed density was very high in the sugar beet trial in this very fertile organic soil (e.g., Chenopodium album, CHEAL, covered almost 50% of the surface in the non-treated control; Supplementary Figures S1B, S2C). Nevertheless, the data show a comparable and respectable efficacy between the ARA's selective and stringent programs and their broadcast-equivalent treatments, with general mean efficacies of 90.3% and 98.8%, respectively (Figure 3A). For now, it is challenging to identify the rationale behind the efficacy gap between ARA selective modality and full spray, as it is difficult to monitor each weed. No difference can be attributed to failed detection and a weak herbicide effect. After three applications of pelargonic acid on sugar beets, the general mean efficacy only reached 60% (Figure 3A). Detailed data showed that three splits of pelargonic acid were sufficient to control Stellaria media (STEME, 98%) but lacked efficacy for other species evaluated in this trial (Supplementary Figure S1B). Despite the high weed pressure, the ARA's non-selective modality managed to lower CHEAL density by 50%. Yet, in conditions of very fertile organic soil and high weed pressure, pelargonic acid does not seem to be an effective standalone approach but should be considered as a complement to a broadcast treatment or a selective spot-spraying program. Nevertheless, pelargonic acid applied using ARA can be considered a weeding solution for sugar beets when the weed pressure is lower. These results show that despite the low mixture volume and molecules applied, ARA controls weeds in both onions and sugar beet crops well, with an efficacy comparable to a broadcast program.

Catch-up treatments and fight against volunteer potatoes

On top of reducing treatment volume with good weeding efficacy, ARA allows for effective catch-up treatments, as exemplified in the onion trial with TE fighting late-germinating AMARE. Targeted ARA sprays eradicated AMARE at BBCH17, with an efficacy >90% after a single herbicide path of a stringent mixture of molecules (Figure 3C, Supplementary Figure S1A, Table 1). This result is impressive, as a broadcast treatment could not have been considered due to the risk to the crop or would have been less effective with a milder mixture.

Similarly, volunteer potatoes in onions can be problematic in some cultural contexts and rotation cycles. Currently, effective herbicide solutions to combat volunteer potatoes in onions are lacking, with no legal molecule available that provides satisfying efficacy. To address this, an ARA spot-spraying treatment was added to the farmer's routine, composed of 18 L/ha of capric acid (288 g/L), caprylic acid (422 g/L) fatty acids, and 0.3 L/ha of fluroxypyr-methyl (480 g/L; Table 3). This program enabled ARA to eradicate 79% of the volunteer potatoes 18 days after treatment (Supplementary Figure S4), with the remaining 21% significantly affected. Notably, no phytotoxicity was observed on the crop. The treatment's success is attributed to using concentrated solutions, the synergistic effect between the three molecules (as a single application of fatty acids or fluroxypyr-methyl did not achieve the 80% efficacy when combined; Supplementary Figure S4B), and a reasonably low density of volunteer potatoes, which allowed efficient weeding without compromising the crop. It is important to note that a higher density of volunteer potatoes might induce collateral phytotoxicity damages on the crop because more product would need to be applied.

In conclusion, ARA controls weeds well and offers new agronomic solutions to emerging problems in which conventional agriculture has gaps.

Although the ARA ultra-precise smart spot sprayer is an efficient weeding solution, ensuring crop health remains a priority. Crop health was monitored throughout the trial, focusing on indicators such as leaf necrosis, plant stunting, thinning, and chlorosis, as defined by European and Mediterranean Plant Protection Organization (EPPO) guidelines (EPPO Standards: General Standards, 2023).

In the onion trials, early phytotoxicity was observed in the broadcast modality, with stunting effects (evaluated at 15% 21 days after TC [21 DATC] and 3% at harvest) and some thinning (17.5%; Figures 3B, 4A, E), indicating a risky program for the farmer (Table 1).

Despite the high concentration of molecules in the ARA selective modality, no stunting phytotoxicity was observed, suggesting that ARA sprays did not affect the crop at early stages. However, ARA selective TD caused mild leaf deformation (5.8%) 7 days after TD (7 DATD), but the onions quickly recovered, with no deformation observed 22 days after TD (Figure 4B). Additionally, pelargonic acid caused minimal necrosis on onion leaves, evaluated at 0.6% 7 days after TD (Figure 4C). This suggests that the erect shape and waxy properties of onion leaves protect the crop from damage caused by sprays targeting nearby weeds.

In the sugar beet trial, no phytotoxicity was observed in the ARA selective modality; however, the same mixture applied in the broadcast modality caused significant stunting, delaying development by 66% compared to the non-treated modality and causing 15% chlorosis at 7 days after TC (DATC; Supplementary Figures S2C, D). Sugar beet plants exhibit planophilic leaves that expose them directly to ARA spot sprays when weeds are close. Unlike onions, sugar beets do not have waxy leaves. Despite this, only 1.6% necrosis was observed in sugar beets 7 DATC, which is negligible 3 weeks post-treatment (data not shown).

Altogether, this suggests that ARA's AI effectively protects crops from herbicide-induced phytotoxicity. ARA enhances the selectivity of herbicides. AI also provides a certain selectivity to non-selective products allowing their use in established crops at high concentrations, as demonstrated with pelargonic acid (680 g/L) at 16 L/ha. This improved selectivity maximizes the action of herbicides against weeds without compromising crop health. ARA thus alleviates restrictions on the available molecules and increases the application concentration thresholds in targeted crops, allowing higher concentrations than in conventional broadcast treatments to be used while reducing the overall number of molecules applied in the field. Consequently, ARA opens new agronomic horizons.

Crop health directly impacts plant biomass production and final yields. As shown earlier, broadcast treatments induced phytotoxicity in both the sugar beet and onion trials from which the crop never fully recovered. To highlight ARA's effects on the crops' plant biomass, raw yield, the number of plants, or individual plant size was quantified. The stunting phytotoxicity observed in the broadcast modality in sugar beets (66%; Supplementary Figures S2C, D) reduced biomass production by half in 10-week-old plants compared to the NT (Supplementary Figure S2E). In the onion trial, the development delay seen early in the broadcast modality (Figures 3B, 4A) persisted until harvest, as shown by delayed senescence (50% of dried leaves) compared to the ARA modalities (Figures 3B, 4H, I). Global onion yields showed significant differences between the ARA modalities and broadcast treatments. ARA selective and non-selective modalities yielded 67 T/ha and 70 T/ha, respectively, while the broadcast treatment yielded 43 T/ha (Figure 4D), suggesting that the broadcast treatment program lowered the yield compared to ARA modalities. Compared to the non-treated modality, ARA increased yields by approximately 20% despite using a stringent program (Figure 4D, Table 1).

ARA's spot-spraying application preserves crop integrity, as shown by the number of bulbs (Figures 4E, I), which were comparable between ARA modalities and the non-treated modality, unlike the broadcast modality, which lost 15% of onions compared to the non-treated modality, directly affecting global yield. Additionally, ARA modalities produced heavier onions compared to non-treated and broadcast modalities (Figure 4F). The size distribution of onions in the ARA modalities was more restricted to the larger group (35–70 mm), representing 93% in ARA selective and 96% in ARA non-selective, compared to 87% in the non-treated and only 80% in the broadcast modality (Figures 4G, I). This restricted distribution in ARA modalities can be explained by the random distribution of weeds in the field, resulting in “random” ARA spray applications that prevent the accumulation of phytotoxicity seen in broadcast treatments and mitigate the negative impact of weeds on plant development. This further supports using the ultra-precise smart spot sprayer in vegetable crops.

Overall, ARA eliminates phytotoxicity, clears plant competition with surrounding weeds, encourages natural crop development, and maximizes crop potential for optimal yield.

The onion trial was led in Switzerland (46.951844, 7.151822), within a farmer's field, on a mineral soil following wheat and lettuce crops. The onions were sown in 1.5-m-wide beds, each containing four rows of onions. The onions were sown with a pneumatic vegetable sowing machine Agricola Italiana SNT-2-290 at the density of 90,000 seeds per hectare, variety: crockett (Bejo, https://www.bejo.com/onion/crockett). Two true-repetition 6 × 30-m plots were designed for the modalities (ARA selective, ARA non-selective, farmer). In each true repetition, two pseudo-replicates were designated. This leads to four pseudo-replicates per modality. The non-treated control represented a 2 × 12-m band separating the ARA replicates (Supplementary Figure S5C). The onions were sown on March 23, 2023, and the crop was monitored up to harvesting on August 18, 2023 (Supplementary Figure S5A).

The volunteer potato trial was led on the same field as the onion trial but on a section where potatoes preceded the onion crop.

The second onion trial presented was led in Germany (48.741404, 12.884032) on silty clay soil, following a winter cover crop composed of white mustard (Sinapis alba) and large-seed false flax (Camelina sativa) sown after wheat crop. The data of each modality represent one single replicate evaluation of a 6 × 50-m plot. The treatment program is available in Supplementary Table S1.

The sugar beets trials were led in Switzerland (46.751874, 6.557217), on our experimental platform, in a loamy highly organic soil, following a corn crop. The sugar beets were sown by the grower with a pneumatic single-grain sowing machine at the density of 117,000 seeds per ha (variety: Agueda, KWS; https://www.kws.com/ch/fr/). Four true repetition plots of 6 × 30 m were designed per modality. In each true replicate, two pseudo-replicates were designated. This leads to eight pseudo-replicates per modality. The non-treated control represented three bands of 2 × 36 m, separating the replicates (Supplementary Figure S5D). The sugar beets were sown on March 23, 2023, and the crop was monitored until June 1, 2023 (Supplementary Figure S5B).

In this study, the quantities of indicated molecules were diluted to an equivalent of 200 L/ha as ARA works on a basis of 200 L/ha of a full spray equivalent. For every treatment applied in this study, the calculated sprayed volumes, and their relative economies were based on 200 L/ha.

The crop programs (calendars and mixtures) can be consulted in Tables 1–3, Supplementary Table S1, Supplementary Figures S5A, B. The volumes/masses of commercially available formulated herbicides were adjusted to the preparation of batches of a final volume of 30 L of mixture and diluted in water under agitation. This 30-L volume is enough to cover the total surface of each modality based on a 200 L/ha (total surface spray). After the initiation process, the mixtures were spread on the field at the speed of 7.2 km/h and with a working pressure of 3.0 bar (3,000 hPa), as recommended by Ecorobotix, to reach an equivalent flow of 200 L/ha. Depending on the algorithm–herbicide combination, a security zone was either set or not (Tables 1–3). To avoid any residues from one modality contaminating the next, the mixture tank was conscientiously rinsed between two modalities. The broadcast treatment performed in the sugar beets corresponds to a 200 L/ha application and, in the onion trial, to a 300 L/ha application.

As the ARA combines a hardware structure controlled by a software, the ARA sprayer is constantly improved and evolves over time (bug fixes, new functions, new crops, etc.). Supplementary Figures S5A, B recapitulated the treatments applied in our trial and the corresponding version releases of ARA over time.

Only herbicide treatments were applied differently between the modalities. All the other crop management interventions were equally applied to the fields and are believed to have not interfered with the results presented here.

The herbicide treatments' efficacy and selectivity were evaluated according to the EPPO standards PP1/152(4), PP1/135 (4), and PP1/181 (5) (EPPO Standards: General Standards, 2023). For onions, the specific PP1/075(3) guidelines were used [PP1/075(3) - Weeds in Allium Crops, 2008]. In the sugar beets trial, PP1/052(4) guidelines were followed [PP1/052(4) - Weeds in Sugar and Fodder Beet and Industrial Chicory, 2020]. To estimate the yield of the onions, consecutive onions of two rows of two meters (1.5 m²) were collected. The onions were immediately calibrated according to the reference table provided by the Swiss association of vegetable market (Prescriptions suisses de la qualité pour les légumes, 2023) in three categories ( ≤ 22 mm, >22–35 mm ≤ , >35–70 mm ≤ ). Each category was weighted, and the number of bulbs per category quantified. Yield was extrapolated to standard unit tons/ha or N/ha. Data show the means of four pseudo-replicates. Twenty sugar beets per repetition were rooted out after 10 weeks' growth and immediately weighed. The data show the mean biomass per plant of the four replicates.

Climatic data were collected by Agrometeo, the Swiss federal climatic service (https://www.agrometeo.ch/). In this study, the provided climatic data correspond to the closest station from the trials: Praz's data were used for the Swiss onion trials, and Pomy's for the sugar beet trials. For the details, see Supplementary Figure S7. For the forecast conditions in the onion trial led in Germany, refer to Supplementary Figure S6E (https://meteostat.net), which shows the Metten station recordings.

Statistical analyses (analysis of variance tests followed by Tukey tests with a confidence level of 95%) were processed on RStudio software (http://www.rstudio.com/) using the Agricolae package (http://cran.r-project.org/web/packages/agricolae/index.html).