Coulomb force-guided deep reinforcement learning for effective and explainable robotic motion planning.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Mobile robot navigation in complex environments presents significant challenges for safety and efficiency.
• Deep Reinforcement Learning (DRL) offers a promising approach for autonomous navigation but often lacks explainability.
• Digital twin technology is increasingly used for training and validating robotic systems.

Purpose of the Study:

• To propose a novel physics-inspired DRL framework for effective and explainable motion planning in mobile robots.
• To enhance robot navigation safety by incorporating attractive and repulsive forces and anticipatory collision avoidance.
• To validate the framework's performance in both simulated and real-world environments.

Main Methods:

• Representing robot, destination, and obstacles as electrical charges, with interactions modeled by Coulomb forces.
• Integrating Coulomb forces into the DRL reward function to guide robot behavior.
• Incorporating LiDAR-based obstacle boundary segmentation for anticipatory collision avoidance rewards.
• Training the DRL model in Gazebo simulations and deploying it on a TurtleBot v3 robot.

Main Results:

• The proposed framework significantly reduces collisions during robot navigation.
• Safe distances from obstacles are consistently maintained.
• The method generates safer and more efficient trajectories toward designated destinations.
• The physics-inspired approach enhances the explainability of motion planning decisions.

Conclusions:

• The physics-inspired DRL framework effectively enables safe and explainable motion planning for mobile robots.
• Coulomb forces and LiDAR-based rewards are crucial for robust navigation in complex, dynamic environments.
• The successful deployment on a real robot validates the framework's practical applicability and performance.