Autonomous Rear Parking via Rapidly Exploring Random-Tree-Based Reinforcement Learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Autonomous rear parking (ARP) is a complex task for nonholonomic vehicles, requiring both efficient path planning and stable path following.
• Existing methods often struggle with real-time performance and adaptability in dynamic parking environments.

Purpose of the Study:

• To propose an efficient and integrated method for autonomous rear parking (ARP) of car-like nonholonomic vehicles.
• To enhance the precision and stability of parking maneuvers through a novel combination of AI and control techniques.

Main Methods:

• Development of an OpenAI Gym environment for training reinforcement learning (RL) agents.
• Implementation of path planning using Rapidly-exploring Random Trees (RRT).
• Integration of Model Predictive Control (MPC) for path following and RL based on the Markov Decision Process (MDP) for decision-making.

Main Results:

• The proposed ARP method demonstrated successful parking maneuvers within the OpenAI Gym environment.
• Evaluation results showed a minimized difference between reference points and generated trajectories, indicating high accuracy.
• The system effectively combined path planning, control, and learning for robust autonomous parking.

Conclusions:

• The proposed integrated ARP method offers an efficient solution for autonomous parking challenges in nonholonomic vehicles.
• The combination of RRT, MPC, and RL provides a powerful framework for achieving precise and stable parking.
• The study validates the effectiveness of the proposed method in a simulated environment, paving the way for real-world applications.