Design and Experimental Validation of Deep Reinforcement Learning-Based Fast Trajectory Planning and Control for...

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Mobile robot navigation in uncertain environments presents challenges in optimal trajectory planning and collision avoidance.
• Existing methods often require extensive training or struggle with dynamic obstacles.

Purpose of the Study:

• To propose a hierarchical deep learning-based control framework for optimal maneuver trajectory planning and waypoint tracking in uncertain environments for mobile robot exploration.
• To enhance the efficiency and performance of autonomous exploration systems.

Main Methods:

• A hierarchical framework with an upper motion planning layer using a recurrent deep neural network (RDNN) for maneuver profile prediction.
• A lower waypoint tracking layer employing deep reinforcement learning (DRL) with a noisy prioritized experience replay (PER) algorithm for collision-free control.
• Leveraging human demonstration data to reduce training time.

Main Results:

• The proposed DRL method demonstrated superior training speed compared to the vanilla PER algorithm in simulations.
• Experimental case studies validated the framework's effectiveness in autonomous exploration.
• The strategy achieved improved motion planning, enhanced collision avoidance, and reduced training time.

Conclusions:

• The hierarchical deep learning-based control framework is effective for autonomous mobile robot exploration in uncertain environments.
• The integration of RDNN for planning and DRL for control offers significant improvements in performance and efficiency.
• The proposed noisy PER algorithm enhances the exploration rate and training speed of the DRL policy.