Una estrategia híbrida de aprendizaje por refuerzo informada por RRT∗ para la planificación de caminos de robots móviles en minería a cielo abierto

This work introduces a hybrid path planning strategy for differential-drive robotic vehicles, combining reinforcement learning methods with sampling techniques. Specifically, Q-Learning (QL) is used to find a global path by exploring and exploiting environmental information, where an agent learns to take actions while maximizing rewards. The agent uses a random sampling method based on Rapidly-exploring Random Trees (RRT∗) to speed up the search of feasible navigation points, combining the advantages of QL with RRT∗ (MQL) to improve sampling and generate smooth and feasible paths in high-dimensional spaces (Smooth Q-Learning - SMQL). The effectiveness of the proposed hybrid method was validated under open-pit mining conditions through a performance analysis based on maneuverability, completeness, reachability, and robustness in environments such as straight roads, narrow spaces, intricate areas, and helicoidal configurations with terrain constraints. Simulations demonstrated that SMQL overcomes the limitations of QL and RRT∗, achieving suitable exploration of the search space and rapid convergence of rewards. Paths previously planned with SMQL and MQL are tested on a motion controller and a Husky A200 robot, achieving a reduction in error cost of 81.9% and 76.4% and control effort of 79.8% and 83.5% compared to QL, respectively. It is expected that these results will impact energy resource savings for the robot when following planned routes in mining environments.