Comparing Multi-Objective Reinforcement Learning approaches for Automowers

Sammanfattning

This study introduces Multi-Objective Reinforcement Learning (MORL) to the Coverage Path Planning (CPP) problem, aiming to balance mowing coverage and safety for autonomous lawn mowers. Three MORL approaches are compared: Predefined Utility Function (PUF), Preference Conditioned Network (PCN), and Multi-Critic Actor Learning (MCAL), all utilizing the Soft Actor-Critic (SAC) algorithm. The agents were trained and evaluated in a simulated Unity environment featuring grass andrisk zones. Quantitative evaluation using the hypervolume metric and qualitative analysis via spatial heatmaps show that the PUF approach performs best overall. PUF successfully achieved the desired trade-offs between safety and coverage across all preference settings. On the other hand, the single-policy methods, PCN and MCAL, struggled to differentiate safer preference settings and exhibited overly cautious behavior, frequently avoiding safe grass near risk zones. Unexpectedly, PCN outperformed the more complex MCAL approach. The results confirm that PUF can effectively manage these trade-offs, while PCN and MCAL may require additional training to fully learn complex safety constraints.