Analysis of Deep Reinforcement Learning Approaches for Autonomous Navigation of Agricultural Unmanned Ground Vehicles in Cassava Farms in Cross River State

Autonomous navigation of agricultural unmanned ground vehicles for crop monitoring and targeted pesticide application in cassava farms offers significant productivity gains for growers in Cross River State, South South Nigeria. However, unstructured and visually complex crop row environments present navigation challenges that classical path planning methods address poorly. This study analyses the performance of three deep reinforcement learning approaches, proximal policy optimization, soft actor-critic, and twin delayed deep deterministic policy gradient, for training autonomous navigation policies for an agricultural UGV in simulated cassava farm environments. A physics-based simulation environment replicating cassava plant morphology, soil surface characteristics, and lighting conditions typical of Cross River State is constructed in Gazebo using ROS. Each DRL agent is trained from a reward function encoding crop row following accuracy, collision avoidance, and energy consumption, and evaluated on held-out navigation scenarios unseen during training. SAC demonstrates the most sample-efficient learning curve, reaching proficient navigation performance after 380,000 simulation steps, compared to PPO at 520,000 and TD3 at 430,000 steps. Sim-to-real transfer of the best-performing SAC policy is evaluated on a physical differential-drive UGV in a 0.4-hectare experimental cassava plot at the University of Calabar Agricultural Research Farm. Real-world row-following accuracy degraded by 14 percent relative to simulation, attributable to unmodelled soil deformation dynamics and variable canopy lighting. The study identifies domain randomization strategies to narrow the sim-to-real gap in future work. Keywords: deep reinforcement learning, autonomous navigation, agricultural robot, cassava farm, Cross River State