Integration of Adaptive Reinforcement Learning Models into the Control of Agricultural Drones

Abstract

This paper presents an integrated approach to adaptive agricultural spraying using unmanned aerial vehicles (UAVs), leveraging reinforcement learning (RL) techniques, particularly the Proximal Policy Optimization (PPO) algorithm. The study focuses on the practical implementation of mathematical models in simulation and onboard control systems. It demonstrates how spray coverage error, chemical loss, and stochastic wind models can be formalized into a reward function and incorporated during RL agent training. The PPO algorithm was implemented using the Stable-Baselines3 library in the AirSim simulator. The agent was trained based on a complex input state vector, including position, wind velocity, crop density, and coverage maps. The training was carried out in stages, starting with low wind conditions and gradually increasing to gusty wind scenarios. The resulting policy was exported in ONNX format and optimized for real-time execution via TensorRT on an NVIDIA Jetson Nano platform, enabling efficient inference onboard the drone. The developed solution was tested in both simulation environments (AirSim, Gazebo) and a physical PX4 SITL platform. A series of experiments were conducted with simulated wind intensities ranging from 2 to 14 m/s. The proposed RL-based adaptive spraying strategy was compared with traditional fixed-parameter control methods. Results showed a reduction in average coverage error by up to 30 % and a decrease in chemical losses by 28 %, confirming the agent’s ability to adapt in real time. A key feature of this approach is its end-to-end practicality: for the first time, a complete development pipeline is presented — from mathematical modeling and training to onboard deployment and real-world validation. The article includes screenshots of the training process, simulated environments, error convergence curves, and the Gazebo GUI, offering transparency and reproducibility for future researchers. This work contributes to the advancement of autonomous precision agriculture systems and lays the groundwork for deploying self-learning UAVs in dynamic field environments.