Curriculum Reinforcement Learning with Hybrid Exploration Strategies for 2D Maze Path Planning

This paper presents a reinforcement learning (RL) framework designed to solve 2D maze path planning problems within simulated environments, with potential applications in microrobot navigation in constrained settings. The primary challenge addressed is the efficient training of an RL agent to navigate mazes ranging from simple 17x17 grids to complex 263x263 grids. Our approach integrates curriculum learning to increase the complexity of mazes during training gradually and employs hybrid exploration strategies to balance exploration and exploitation effectively. Although the experiments were conducted in a simulated environment, the methods developed hold significant promise for real-world microrobot applications. We benchmarked our RL framework, which is built on a Deep Q-Network (DQN), against traditional pathfinding algorithms such as Breadth-First Search (BFS), A-star, and Rapidly-exploring Random Trees (RRT). We trained our model on the subset of the ‘Curriculum - Lost in Maze Dataset,’ consisting of 3.1 million maze images, demonstrating the model's ability to generalize and efficiently solve unseen mazes. The model achieved an average loss of 12.67 across all maze sizes and adapted to new ones with an average time of 21.69. The proposed methodology results in significant improvements, with a 16.54% increase in success rate and a 51.49% improvement in path efficiency compared to the average of the two baseline methods. Our proposed approach potentially contributes to microrobot manipulation, where precise path planning is crucial for avoiding obstacles in real-time within a vision-based control system limited to a monocular top view by the optical microscope [1], [2], [3].