Informed Reinforcement Learning for mobile robot planning and control tasks based on the D* algorithm

Abstract

Autonomous navigation of mobile robots in partially unknown and dynamic environments requires control strategies that are both optimal and robust to disturbances.This paper proposes a hybrid navigation framework that integrates global path planning, Pontryagin-based optimal control, and informed reinforcement learning. The system is implemented as a hierarchical ROS2-based architecture, where D*-Lite generates a global path, an optimal control module computes nominal control actions, and a TD3-based RL agent learns corrective policies to compensate for disturbances.

Unlike conventional RL approaches, the proposed method leverages optimal control solutions as informative priors, allowing the agent to learn residual control policies instead of learning from scratch. This significantly reduces the exploration space and improves training efficiency.

Experimental results in a Gazebo simulation with dynamic obstacles and physical perturbations demonstrate improved stability, faster convergence, and better trajectory tracking compared to standalone methods.

The proposed approach provides a scalable and robust framework for integrating model-based control and data-driven learning in mobile robot navigation.