Comparing and analysing the effectiveness of Multi-Agent Reinforcement Learning (MARL) algorithms for simplistic coordination in row cultivation applications

Precision agriculture (PA) is revolutionizing farm management through AI, robotics, and sensors to optimize resources, yield, and sustainability, yet lacks simple MARL benchmarks for cooperative crop inspection under partial observability and multi-objective rewards. This study implements a standardized PettingZoo/Gymnasium environment where two homogeneous agents traverse a 32-plant row over 5000 episodes for binary disease detection, systematically comparing PPO, DQN, MADDPG, and QMIX using discrete actions (move/inspect) governed by a reward scheme balancing true detections (+5), false positives (-0.1), repeats (-0.2), and movement efficiency (+0.1 forward, -0.01 backward). PPO achieved superior performance with stable episodic rewards of 0.7, 5.27 hits per episode, and near-zero false positives/repeats, outperforming value-based and CTDE methods that stagnated or collapsed. These results establish PPO as a robust, modular baseline for scalable crop monitoring deployable in VRT, yield surveillance, and real-world autonomous scouting requiring continuous learning.