Development of a Weeding Agrobot Prototype for Inter-row and Intra-row WeedingDevelopment of a Weeding Agrobot Prototype for Inter-row and Intra-row Weeding

There is a critical farm labour shortage and limited machinery usage in India. Weeding removes undesirable plants from fields, which compete with crops for living, reducing yield and crop quality. Inter-culture or weeding operations are not fully mechanized. Mechanical weeding is preferred over manual and animal-drawn methods, which are labour-intensive. In India, available weeders mainly remove inter-row weeds, which are costly and not autonomous. Oversized machines are unsuitable for average farm sizes and cropping patterns. In recent innovations, laser technology or weedicide sprayers have been integrated into weeders. An intra-row weeding system was developed based on mechatronic techniques (Kumar et al., 2024). The performance of a self-propelled diesel engine equipped with three-blade sets, designed for intra-row weeding in vegetable fields, was evaluated (Khayer et al., 2024). The performance of the power weeder was assessed in sugarcane crops under various conditions (Mohan et al., 2020). An e-powered mechanical inter-row weeder has been developed for small farms (Pandey et al., 2023). The performance of a battery-operated power weeder was assessed in a soybean crop (Pandey et al., 2021). The integrative weed management treatments on growth attributes were evaluated (Bandyopadhyay et al., 2024) and effect on rice (Kashyap et al., 2022) and overall productivity of crops was studied (Meena et al., 2022). A lightweight weed detection mechanism for laser-weeding robots was developed using crop-weed datasets, achieving a mean Average Precision of 0.88 and 0.53 at an Intersection over Union (IoU) of 0.5 with YOLO5 and SSD-ResNet50 models, respectively (Fatima et al., 2023). RetinaNet and YOLOv5n reached mAP of 79.98% and 76.58% in the cotton field,  as mentioned by (Rahman et al., 2023).
       
Due to the gaps, such as labour dependency, unsuitable size and high cost of the available weeders and unavailability of intra-row weeders, the objective was to develop a compact, low-cost, efficient, autonomous, environmentally friendly weeder for inter- and intra-row weeding to assist small farmers. The research objectives include identifying gaps in the present study, developing a prototype of an autonomous weeder for inter-row and intra-row operations and testing the prototype in a sample soil test bin to validate the concept.

The present work was carried out at the Department of Instrumentation and Control Engineering, COEP Technological University, Pune, Maharashtra, India, for period of three years from 2022 to 2025. Considering the Indian farming conditions, a prototype of an autonomous weeder is developed through the steps mentioned as follows.
 
Dimensional design of an autonomous weeder prototype
 
Fig 1 (a) represents the dimensional scenario of the Agrobot, along with the assumed dimensions of the inter-row and intra-row spaces, 300 mm and 150 mm, respectively Fig 1 (b) shows the actual onion field on the test soil bin and Fig 1 (c) represents an actual prototype.


 
Motor selection
 
The total force is the summation of the rolling resistance force, aerodynamic force and acceleration force. Assuming the motor efficiency is 85%, the power required is 0.4 kW for muddy and 0.6 kW for sandy soil. An electric brushless direct current  BLDC motor with the criteria 0.6 kW < Motor power and 350 < Motor rpm was selected to move the weeder.

Design in CATIA and stress analysis in ANSYS
 
The mechanical frame structure was designed using CATIA software and stress analysis was done using ANSYS software (Fig 2a, 2b). When 250 N was applied, the maximum stress and deformation observed were 7.8 MPa and 0.11 mm, respectively.


 
Components of the weeder
 
1. Frame: Made of mild steel (MS), affordable, weldable and machinable.
2. Four wheels: The front wheels are bike tyres and the rear wheels are small activa tyres. The motor is attached to the front wheels through a chain drive-pull and conveyor belt.
3. Web camera: Zebronics zeb-crisp pro web camera (HD), mounted vertically on the frame to create a database for machine learning.
4. Pneumatic cylinder, actuator: The pneumatic cylinder and actuator, controlled by a Raspberry Pi, operate on 24V and 12V, respectively, to control a centrally connected weeding tool.
5. Battery: 48V (Four 12V lead-acid rechargeable batteries in series)
6. Raspberry Pi 4B+: This device has a 64-bit quad-core processor, 8GB of RAM, moderate power consumption, high processor speed, multimedia performance, memory and connectivity.
7. Inter-row, intra-row weeding tools: Two centrally located intra-row weeding tools and two both-sided inter-row weeding tools (HO-M15 Bow, WOLF-GARTEN tools) are market-available.
 
Dataset collection, data preparation and building a machine learning model
 
A spring onion farm was cultivated to prove the concept and conduct trials. After many object detection models were tried, the best model was chosen.
 
Spring onion cultivation
 
Spring onions are vulnerable to weeds due to their gradual early development and lack of adequate foliage to cover the ground, which hinders weed growth. The spring onions were grown in the designated test soil bin in the University area, with an inter-row spacing of 300 mm and an intra-row spacing of 150 mm. The plants grown on the farm other than the onion crop were assumed to be weeds. (Table 1) shows the details of the test soil bin.


 
Dataset preparation
 
RGB images of weeds and spring onions were collected using a Python program developed on Raspberry Pi and a USB web camera positioned 150 mm above the ground while traversing the soil bin. The web camera captured images of 640 x 480 pixels. Photographs of 2600 × 4624 pixels, taken with a mobile phone camera, were added to enhance the dataset’s diversity. All photographs were taken on sunny, cloudy and windy days, in dry and wet soil conditions, encompassing the growth stages from planting to harvesting and saved in .jpg format (Fig 3a-3d). Out of 5,000 images, including blurred photos, duplicates and those without onions or weeds, were removed for better accuracy.


 
Object detection models
 
Convolutional neural networks (CNN) consist of input, convolutional, pooling and flattened layers and they use tensors as the foundation for object detection. The transfer learning leverages weights from pre-trained models to improve performance with minimal custom data. LabelImg software was utilized for image annotation. The dataset was split into training, testing and validation sets in ratios of 80%, 10% and 10%, respectively. A balanced dataset was used for two classes: onion and weed. EfficientDetD0, MobileNetV2 SSD, MobileNetV2 SSD FPN-Lite, YOLOv8s and YOLOv8n were trained through varying epochs.
 
Development of a python code and deployment of a Machine learning model on raspberry Pi
 
The following functionalities are added by preparing the code in Python on the Raspberry Pi:

1. PWM (Pulse width modulation) signal generation for motor control: GPIO pin 18 was used as the output to control the 48V DC motor by generating a pulse width modulation (PWM) signal. The PWM duty cycle is used to vary the motor speed by controlling the motor terminal voltage (Katsambe et al., 2017).

 

Where,
t_ON = ‘ON’ time.
T = Duration of one period = t_ON + t_OFF.
V_in = Voltage supplied to the motor.
V_out = Voltage after duty cycle is applied.
 
2. Program for controlling a pneumatic cylinder and actuator through relays (for controlling an intra-row weeding tool): GPIO pins 21 and 26 were used as outputs from the Raspberry to input a control signal to two relays that control the pneumatic compressor and actuator, which operate at 12V and 24V, respectively.
3. Image capturing: One program was deployed on the Raspberry Pi to capture and save pictures every second from a continuous video stream captured by the webcam.
4. Deployment of best-suited machine learning model on Raspberry Pi: Google Colab facilitates the training, conversion and export of TensorFlow Lite and PyTorch models.
 
Evaluation of the system
 
Evaluation of object detection models
 
Precision refers to the proportion of accurate predictions relative to the total number of predictions made. Recall signifies the ratio of correct predictions to the number of relevant objects present. Average precision (AP) is the area beneath the precision-recall curve, illustrating the trade-off between precision and recall. In mAP, the predicted bounding boxes from the model are compared against the ground truth annotations. The mean of the average precision (AP) across all classes and various Intersections over Union (IoU) thresholds is taken to compute it.
 
Evaluation of the proposed weeder
 
The following are the evaluation parameters of the weeder that involve evaluating its quality and quantity of work and the power used:
 
1. Width of the operation: The total width of inter and intra-row weeding tools.
2. Depth of operation: The vertical distance measured from the top of the soil surface to the base of the excavated area using a steel scale.
3. Speed of the weeder: The ratio of distance travelled by the weeder to the time required.

 

Adjusting the code’s PWM duty cycle can set the weeder’s speed. The speed was calculated using a stopwatch, considering the time required to travel a 10 m distance.

4. Theoretical field capacity (TFC in ha hr^-1): The machine’s theoretical field capacity refers to how much ground it can cover, assuming it operates at full speed and utilizes its entire rated width without downtime (Devojee et al., 2019).

 

5. Actual or effective field capacity (EFC in ha hr^-1): The area weeded during each trial run within a specified time interval was recorded to calculate the weeder’s Effective field capacity using a stopwatch (Devojee et al., 2019).

 

6. Field efficiency (%): Field efficiency is the ratio of effective field capacity to theoretical field capacity.

 

7. Weeding efficiency (%): The ratio between the number of weeds removed by the prototype weeding agrobot and the number of weeds in a unit area, expressed as a percentage. The average of the three trials was taken.

 

W₁ and W₂ are the number of weeds counted in a specified area before and after the weeding operation. The average of inter-row and intra-row weeding efficiency is used to calculate weeding efficiency.

8. Plant damage (D%): The ratio of the number of plants/ crops damaged after the weeding operation in a row to the number of plants present before the operation is calculated as follows:

 

P₁ and  P₂ are the number of crops counted in a 10 m row field length before and after the weeding operation. The plant damage primarily arises from insufficient crop spacing, delays between weed detection and the operation of the weeding tool and inaccuracies in the machine learning model. Plant survival is calculated as follows:

 
 9. Performance index (PI): It is directly related to the field capacity (ha/hr), plant survival and weeding efficiency (%) and is inversely associated with power (hp) input (Balas et al., 2023).

 

Where,
EFC = Effective field capacity (ha hr-1).
WE = Weeding efficiency (%).
D = Plant damage (%).
P = Power used in hp.
       
Power input to the weeder can be calculated as follows: (Sahu et al., 2017).

 

Where the draft can be calculated as follows:

 

The black soil’s specific resistance is 0.4 kg cm^-2 (3.924 N cm^-2).
 
Working flow
 
When the switch is ON, the motor starts running, so the weeder begins moving. Inter-row weeding blades on both sides will remove inter-row weeds as the weeder moves. Simultaneously, the camera mounted on the weeder will capture photos of the intra-row space. The machine learning model deployed on the microcontroller will differentiate the main crop from weeds. Until the weeds are detected by the machine learning model deployed on Raspberry Pi, the weeder will not stop. Suppose the image captured is identified as a weed. In that case, the Raspberry Pi will first stop the motor through a PWM signal and activate the relays connected to the pneumatic cylinder and actuator. It will bring the intra-row weeding tool down; after a few seconds of delay, the motor will restart, allowing the weeds to be removed through a forceful dragging action. The intra-row weeding tool will be backed up by actuator and the motor will continue to move. Fig 4 illustrates the workflow of a Weeding Agrobot Prototype.

Results of object detection models
 
Out of the object detection models tested, YOLOv8n yielded the best results. The model size is tiny, 5.96 MB, making it work properly on the Raspberry Pi. The inference time for the YOLOv8n model is 2ms. The mAP IoU@0.5 of 0.85 for YOLOv8n (trained for 50 epochs) gave good prediction accuracy. Therefore, this model would be suitable for autonomous weeders to detect weeds accurately. Table (2) shows the model size-wise and mAP-wise comparison of the different tried models.

​
       
Different trials of YOLOv8n models are mentioned in Table (3).

​
       
The YOLOv8n model was trained for different epochs, yielding better accuracy at 50 epochs. Its performance was tested using both test data and actual data from the onion farm. Fig 5 (a), (b) show the YOLOv8n model’s prediction results on test data and the actual field when working in real-time with an autonomous weeder, respectively. Fig 5 (c), (d) Show the normalized confusion matrix for the YOLOv8n model and the precision-recall (PR) curve for the test data, respectively. The mAP@50% threshold was 0.87 and 0.82 for onion and weed detection, respectively.


 
Performance evaluation of the prototype weeding agrobot
 
Table 4 shows the parameters of the Weeding Agrobot Prototype and the (Table 5) shows the weeder’s average speed, weeding efficiency and plant damage.

​

​
       
The weeding operation costs approximately INR 80,000 per hectare of land for the sugarcane crop (Mohan et al., 2020). For the onion crop, this cost increases to INR 12,000. The operation cost of the developed prototype, costing INR 80,000, was calculated by referring to the study presented by author Mohan et al., (2020). The hourly fixed cost works out to be INR 46.67. The hourly variable cost, which includes charging costs, repair and maintenance, amounts to INR 11.5. Considering the field capacity, the hourly operation cost per hectare would be INR 2,529. That means the prototype significantly reduces weeding operation costs by 80% compared to traditional methods.
 
Experimentation
 
PWM was set at different values and the performance index for every setting was calculated. At 40% duty cycle, the best results were achieved (Table 6). The intra-row space was varied and the intra-row weeding efficiency was evaluated. For a very small intra-row distance, the machine learning model did not perform well; therefore, various experiments were conducted to explore different widths of the intra-row spaces. At an intra-row space of 150 mm, the weeder’s performance index was the best (Table 7). Table 8 summarizes the models’ performances from previous work and compares them with the proposed best-suited YOLOv8n model. As the datasets differ for different crops, directly comparing the proposed model with other previously used models is invalid. The YOLOv8n, nano version model, suitable for real-time applications, gave comparable precision and recall values for the difficult onion-weed detection dataset. Table (9) compares the proposed weeder’s different parameters with those of other weeders in the literature. The developed weeder is an autonomous weeder that operates on electric energy, suitable for both inter-row and intra-row operations. In contrast, other weeders either run on fuel or only remove inter-row weeds.

​

​

​

​

The study concluded that the developed prototype of the Weeding Agrobot for inter and intra-row weeding is compact, easily movable, cost-effective and suitable for small farms in India. To our knowledge, there have been very few previous attempts to classify spring onions and weeds in onion fields. This research contributed a diverse dataset of spring onions and weeds. YOLOv8n was evaluated on the test data and run in real-time by deploying it on a Raspberry Pi, which was used in an Autonomous weeder moving at a speed of 0.12 m/s to cut intra-row weeds within a 150 mm space. The mAP@50% of 0.85 was achieved with an average weeding efficiency of 86% and plant damage of 4%, resulting in a performance index of 1726.26 in real-time applications. Future developments may enable the collection of vast amounts of data, adjustable wheel distances per row, automatic steering, more efficient weeding tools, specialized agricultural tyres and solar-powered batteries. The presented work will guide further work in providing a machine vision to the autonomous weeder and automatically removing weeds. This will help address the labour scarcity problem in agriculture and benefit farmers by enhancing weed management in India.

The present study was supported by the Centre of Excellence in Industrial Automation, COEP Technological University, Pune, Maharashtra, India, and also provided space for the experimental trials on the University campus.
 
Disclaimers
 
The views and conclusions expressed in this article are solely those of the authors and do not necessarily represent the views of their affiliated institutions. The authors are responsible for the accuracy and completeness of the information provided, but do not accept any liability for any direct or indirect losses resulting from the use of this content.

The authors declare that there are no conflicts of interest regarding the publication of this article. No funding or sponsorship influenced the design of the study, data collection, analysis, decision to publish or preparation of the manuscript.