Simulation-Based Autonomous Lawn Mowing Using Double Deep Q-Network

This paper presents a RL approach to developing an autonomous lawn mower agent that can perform efficient mowing of lawns and learn to adapt to changing environmental conditions. The agent is deployed in a grids based simulation environment with various grass heights, static obstacles, and random weather conditions that affect grass growth rate. Training is performed using a Double Deep Q-Network (DDQN) algorithm, which has been adapted to reduce the overestimation bias conventionally found with the baseline Deep Q-Network (DQN) algorithm. Performance evaluation using metrics of total reward, collision avoidance rate, coverage rate, and efficiency ratio shows that the DDQN agent performs significantly better than DQN and rule-based benchmark techniques. Experimentation outcomes reveal that the agent acquires optimal policies for maximum lawn coverage with minimal collision and shows clearly distinct adaptive strategies under changing weather conditions. The results affirm the effectiveness of deep reinforcement learning, specifically the DDQN approach, with an effectiveness of performance at 94.92% in adaptive control-policy development for autonomous agents in a simulated environment.