A quadruped robot for real-time pest detection in agriculture using GAN-enhanced deep learning

In recent years, technological solutions have become increasingly important in agriculture to increase productivity and minimize environmental impacts. This paper presents a quadruped robot system for the detection and classification of pests in agriculture using generative artificial intelligence (AI) and autonomous robotic systems. The system enriches the dataset by generating synthetic data using generative adversarial networks (GANs) based approach and thus achieves high accuracy rates despite the limited amount of real data. Using the deep convolutional GAN (DCGAN) algorithm, 2000 synthetic insect images were generated from 400 real insect images (leaving 100 original images completely isolated as an unseen test set). Using these images, deep learning models such as DenseNet-121 and ResNet-152 were trained. The results show that when trained on the augmented dataset with DCGAN, the ResNet-152 model achieved the highest success with 94% accuracy and an F1 score of 0.93, significantly increasing the 76.4% accuracy obtained using only real images. The Gazebo and RViz simulations confirmed the system’s resilience against physical restrictions, such as gravity (9.81 m/s 2 ), a friction coefficient ( μ = 0.6 ) and fluctuating wind speeds (0–5 m/s). The crawl gait algorithm maintained the robot’s balance by keeping its center of mass within the support polygon, thereby enabling it to move smoothly over rough terrain. Field prototype tests confirmed the functional feasibility of the integrated system for autonomous and precision pest management in sustainable agriculture.