A Mapless Local Path Planning Approach Using Deep Reinforcement Learning Framework.

Area of Science:

• Robotics
• Artificial Intelligence
• Machine Learning

Background:

• Autonomous mobile robots require robust path planning and obstacle avoidance, especially in dynamic, unknown environments.
• Current local path planning methods struggle with unpredictability and limited environmental data.

Purpose of the Study:

• To develop an end-to-end local path planner for autonomous mobile robots using deep reinforcement learning.
• To address challenges in dynamic environments by improving exploration and reward mechanisms.

Main Methods:

• Proposed RND3QN (n-step dueling double DQN with reward-based ϵ-greedy) algorithm utilizing LiDAR data.
• Implemented n-step bootstrapping to reduce bias and a reward-based ϵ-greedy strategy for enhanced exploration.
• Introduced an auxiliary reward function to address sparse reward issues.

Main Results:

• RND3QN achieved higher average total reward values compared to D3QN in simulation.
• Demonstrated significant success rate increases of 174%, 65%, and 61% over D3QN across three stages.
• Validated effective transfer of learned strategies from simulation to a real TurtleBot3 waffle pi robot.

Conclusions:

• RND3QN offers a superior solution for local path planning in challenging, dynamic environments for autonomous robots.
• The algorithm's effectiveness is proven through simulation and real-world robot experiments.
• The developed strategies show strong generalization capabilities for real-world deployment.