Deep reinforcement learning and robust SLAM based robotic control algorithm for self-driving path optimization.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Robotic control during maneuvers often suffers from low accuracy and poor robustness in target path tracking.
• Existing Simultaneous Localization and Mapping (SLAM) algorithms struggle with robustness and dynamic object interference in visual sensing.

Purpose of the Study:

• To propose a novel reward shaping deep deterministic policy gradient (RS-DDPG) algorithm for enhanced robotic path tracking.
• To develop a robust visual SLAM algorithm for dynamic scenes using semantic segmentation and geometric information.

Main Methods:

• Implemented a reward shaping mechanism within the Deep Deterministic Policy Gradient (DDPG) framework to optimize tracking parameters.
• Developed a visual SLAM algorithm integrating semantic segmentation and geometric information for improved dynamic scene handling.
• Utilized the Apollo autonomous driving simulation platform for comparative performance analysis.

Main Results:

• The RS-DDPG algorithm demonstrated superior path tracking accuracy and robustness compared to the standard DDPG algorithm.
• The enhanced visual SLAM system showed significantly improved performance in dynamic scenarios.
• Simulation experiments validated the effectiveness of the proposed integrated approach.

Conclusions:

• The proposed RS-DDPG algorithm offers a significant improvement for robotic path tracking accuracy and stability.
• The integration of semantic segmentation and geometric information enhances the robustness of visual SLAM in dynamic environments.
• This research contributes to more reliable autonomous systems in complex, real-world conditions.