N-DQN-RRT: A Hybrid Deep Q-Network Guided RRT Algorithm for an Efficient Path Planning in Two-Dimensional Environments

Abstract

Autonomous navigation requires path-planning algorithms that can find safe and efficient routes while avoiding obstacles. Classical sampling-based planners such as Rapidly-exploring Random Tree (RRT) and RRT* are widely used because they are flexible and easy to apply in different environments. However, these methods often produce longer paths and require many samples in complex spaces. This paper proposes Neural Deep-Q Network RRT (N-DQN-RRT), a hybrid path-planning algorithm that integrates a Deep Q-Network (DQN) agent into an RRT*-style planning framework. The proposed method uses a directed-random sampling strategy. In this strategy, 75% of samples are guided by the trained DQN policy, 5% are assigned to goal bias, and 20% are sampled uniformly at random. The random component is retained to reduce premature policy bias and maintain exploration. The planner also uses an adaptive rewiring radius that changes according to local obstacle conditions. This allows the method to improve local path quality while avoiding unnecessary neighbor searches in cluttered regions. The algorithm was evaluated in three static two-dimensional environments with increasing map sizes of 10 × 10, 25 × 25, and 50 × 50. Each environment used randomized obstacle layouts and 100 independent trials, giving 300 trials per algorithm. The proposed method was compared with RRT and RRT* using six metrics: path length, steering angle, iteration count, sampling size, computation time, and success rate. Results were reported as mean ± standard deviation with 95% confidence intervals. N-DQN-RRT achieved a 100% success rate across all 300 trials, while RRT achieved 10–100% and RRT* achieved 14–100%, depending on map complexity. In the largest environment, N-DQN-RRT reduced path length by 11.9%, steering angle by 49.5%, sampling size by 75.1%, and computation time by 91.9%compared with RRT*. These results show that DQN-guided sampling can improve the empirical efficiency of RRT*-style planning in static two-dimensional obstacle environments.