EnviroBot: Towards a Transfer Learning and Computer Vision-based Herbicide Applicator Robot

The proliferation of broadleaf weeds on lawns is a prevalent issue in Australia, impacting the aesthetic appeal of lawns and obstructing walking areas. Traditional manual herbicide spraying poses risks to human health, while indiscriminate herbicide application results in environmental pollution and wastage. To address this, we developed EnviroBot, a transfer learning and computer vision-based herbicide applicator prototype, to identify and locate Plantain weeds in lawns accurately. Our prototype utilised a web camera, illuminator, and shade cover to capture training images annotated and used for training with YOLOv8’s nano model. Transfer learning was employed to fine-tune the YOLOv8 model, resulting in an overall precision of $78 \%$ after 391 epochs of training. Live testing demonstrated the model’s effectiveness in identifying and localising Plantain weeds precisely and suitable for real-time usage. Future improvements could involve employing multispectral cameras and expanding the training dataset.