Performance Evaluation of Move Base Integration with Deep Reinforcement Learning for Autonomous Robot Navigation*

The integration of global path planning strategies in reinforcement learning based robot navigation remains an open question in robotic research. This study evaluates the impact of incorporating a global strategy —namely the strategy of Move Base from ROS—on the effectiveness of a robot using a Twin Delayed Deep Deterministic Policy Gradient (TD3) model for navigation. The primary objective is to determine whether Move Base enhances performance in terms of success rate, collision rate, navigation time, and distance traveled when combined with TD3-based local planning. The experimental evaluation is conducted using a previously trained TD3-based model within the Noetic ROS framework, integrating a custom OpenAI Gym environment and the Gazebo simulator to test two setups: (1) TD3 with Move Base (global strategy) and (2) TD3 without Move Base (non-global strategy using the competition package). Results indicate that excluding Move Base leads to improved navigation efficiency, achieving higher success rates, lower travel times and shorter distances in both simple and complex environments. Although Move Base provides a slight advantage in reducing collisions in dynamic environments, its overall contribution to efficiency is limited. These findings highlight the need for careful consideration when integrating global path planning strategies with reinforcement learning based navigation.