Robust Path Planning via Deep Reinforcement Learning

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Deep reinforcement learning (DRL) for robot navigation suffers from inconsistent performance due to stochastic actions and slow learning from inefficient exploration.
• Traditional path planning offers stability but lacks flexibility in dynamic environments.

Purpose of the Study:

• To enhance the robustness and efficiency of autonomous mobile robot path planning.
• To address limitations in DRL navigation, including performance inconsistency and delayed learning.
• To integrate the strengths of DRL and traditional path planning for improved navigation.

Main Methods:

• A hybrid approach combining Twin Delayed Deep Deterministic Policy Gradient (TD3) with classical path planning.
• Pre-processing LiDAR data to extract essential features for state representation, improving computational efficiency.
• Prioritizing negative reward experiences during training and dynamically comparing TD3 actions with classical path-planning actions.

Main Results:

• The proposed hybrid model achieved significantly higher and more stable reward values compared to existing DRL networks.
• The approach demonstrated robust improvement in path-planning processes, validated through simulation experiments.
• Averaged results over 10 independent runs confirmed statistical significance and mitigated random initialization impact.

Conclusions:

• The hybrid DRL and classical path-planning approach significantly enhances autonomous mobile robot navigation robustness.
• Pre-processing sensor data and optimizing the learning process lead to more efficient and stable path planning.
• This research offers a promising direction for developing more reliable autonomous navigation systems.