Reinforcement learning-based dynamic obstacle avoidance and integration of path planning.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Deep reinforcement learning enables complex behavior learning from data.
• Decentralized systems offer advantages in scalability and robustness for multi-agent scenarios.

Purpose of the Study:

• To develop a decentralized collision avoidance framework for mobile robots using reinforcement learning.
• To enhance navigation efficiency by integrating path planning with obstacle avoidance.
• To validate the proposed method in dynamic and complex environments.

Main Methods:

• A decentralized reinforcement learning framework was proposed for collision avoidance.
• The Soft Actor-Critic algorithm was employed for training obstacle avoidance policies.
• A path planner was integrated to ensure path efficiency and avoid dead ends.
• The trained policy was implemented using the Robot Operating System (ROS).

Main Results:

• Mobile robot agents successfully learned to avoid dynamic obstacles and reach target points efficiently.
• The integrated path planner addressed situations where traditional pathfinding might fail.
• The system demonstrated effectiveness in both simulated and real-world environments with differential drive robots.

Conclusions:

• The proposed decentralized reinforcement learning framework provides an effective solution for mobile robot collision avoidance in dynamic environments.
• Integration with path planning enhances navigation reliability and efficiency.
• The method is validated for practical application in robotics.