FL-GQL: a fuzzy logic-guided Q-learning algorithm for global path planning of mobile robots in grid environments

Area of Science:

• Robotics and Artificial Intelligence
• Reinforcement Learning
• Path Planning Algorithms

Background:

• Traditional Q-learning struggles with slow convergence and inefficient exploration in grid-based path planning.
• Existing methods often focus on single Q-learning components, limiting overall performance improvements.

Purpose of the Study:

• To develop a novel fuzzy logic-guided Q-learning algorithm (FL-GQL) for enhanced global path planning.
• To address limitations in convergence speed, path quality, and planning reliability of existing methods.

Main Methods:

• Integration of heuristic Q-table initialization, hybrid reward shaping, and fuzzy-guided dynamic exploration.
• A unified learning framework coordinating multiple Q-learning components.
• Comparative experiments across diverse grid environments with varying complexities.

Main Results:

• FL-GQL achieved a 66.43% success rate, outperforming BQL (37.16%) and DQN (43.87%).
• Demonstrated improved convergence efficiency, path smoothness (fewer turns), and planning reliability.
• Ablation studies confirmed the synergistic contribution of FL-GQL's integrated components.

Conclusions:

• FL-GQL offers an effective solution for global path planning of autonomous mobile robots in structured grid environments.
• The coordinated approach of initialization, exploration, and reward shaping is crucial for performance gains.
• FL-GQL shows robustness and stable performance across varied parameter settings.