Fusion of an A-RRT planner and a supervised model for reactive and optimal navigation of mobile robots

In dynamic and partially unknown environments, ensuring safe and efficient navigation for mobile robots remains a major challenge. This article proposes an improved implementation of a hybrid trajectory planning framework that combines a globally optimal trajectory planner with a reactive local planner to enhance navigation robustness. The global planner is based on an improved Rapidly-exploring Random Tree that incorporates a selective sampling strategy inspired by the heuristic function of the A* algorithm, significantly reducing planning time while improving path quality and smoothness. For local adaptation, a supervised learning model using Extreme Gradient Boosting is trained on a large dataset of robot-environment interactions to predict safe navigation actions in real time. To increase system reliability, a decision correction module is integrated to filter out risky or repetitive actions. A dynamic fusion mechanism ensures seamless coordination between the global and local planners, allowing the robot to adapt effectively to static, unknown, and mobile obstacles. Extensive simulations validate the proposed method. The global planner an improved Optimal Rapidly-exploring Random Tree algorithm reduces computation time by up to 68% compared to the A* algorithm and 91% against the standard Optimal Rapidly-exploring Random Tree algorithm in complex environments while maintaining near-optimal path lengths. The local planner, leveraging an Extreme Gradient Boosting model with 82% predictive accuracy, ensures high responsiveness. This fusion demonstrates superiority in trajectory optimality, safety, and computational efficiency compared to conventional planners.