Path-Tracking Control Strategy of Unmanned Vehicle Based on DDPG Algorithm.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Autonomous vehicles require robust path-tracking capabilities for safe navigation.
• Existing algorithms like DDPG face challenges with Q-value overestimation and slow training.
• Effective path tracking is crucial for unmanned vehicle operation in diverse scenarios.

Purpose of the Study:

• To develop and evaluate a deep reinforcement learning-based path-tracking controller for unmanned vehicles.
• To address the limitations of the Deep Deterministic Policy Gradient (DDPG) algorithm in terms of training speed and Q-value estimation.
• To enable autonomous learning of path-tracking skills through interaction with a realistic simulation environment.

Main Methods:

• Proposed a novel controller using the Double Critic Network Deep Deterministic Policy Gradient (DCN-DDPG) algorithm.
• Formulated a Markov decision process model, defining state, action, and reward functions.
• Trained the control strategy using offline learning within the CARLA simulator's Town04 urban environment.

Main Results:

• The DCN-DDPG controller achieved successful path tracking under various conditions in a simulated urban setting.
• Performance was benchmarked against the original DDPG algorithm, Model Predictive Control (MPC), and Pure Pursuit.
• The proposed strategy demonstrated significant improvements in tracking accuracy and efficiency.

Conclusions:

• The DCN-DDPG algorithm offers enhanced environmental and speed adaptability for unmanned vehicle path tracking.
• The developed controller provides a viable and effective solution for autonomous navigation challenges.
• This research validates the efficacy of deep reinforcement learning for advanced vehicle control applications.