Optimized Path Planning using Double Deep Q-Networks with Dynamic Window Approach

Autonomous mobile robots require path-planning methods that solve complex static worlds with strong safety and efficiency guarantees. Standard Deep Q-Network (DQN) planners are typically unstable and subject to Q-value overestimation, leading to collision or suboptimal paths. In this paper, a hybrid planner is presented that uses the Dynamic Window Approach (DWA) to supply kinematically valid, collision-checked trajectory proposals to a Double Deep Q-Network (DDQN) agent for stable high-level trajectory choice. In two static simulation worlds, a simple layout, and a sparsely crowded layout, the hybrid planner achieves much higher success rates and zero collisions compared to a standalone DQN and DDQN. In the crowded world, it achieves an 88.3% goal-reach rate with zero collisions, compared to 68.6% success rate and a 27.7% collision rate for standalone DDQN. Training time in complex worlds increases from 9.17 hours (standalone DDQN) to 19.70 hours (hybrid DDQN+DWA), being more conservative in exploration. The results indicate encouraging potential for handling dynamic obstacles and implementation in the real world.