SmartGridDrive: an integrated adaptive Q-learning framework for precision self-parking and reverse navigation in...

Area of Science:

• Robotics and Autonomous Systems
• Artificial Intelligence and Machine Learning

Background:

• Autonomous vehicles require robust navigation and parking in complex, dynamic environments.
• Existing methods often struggle with real-time adaptation to moving obstacles and changing parking availability.

Purpose of the Study:

• To develop an adaptive Q-learning framework for precise self-parking and reverse maneuvering.
• To enhance vehicle autonomy in dynamic, grid-based environments with unpredictable elements.

Main Methods:

• An adaptive Q-learning approach with epsilon-greedy parameter decay based on rewards.
• Look-ahead obstacle prediction utilizing a Markov transition model.
• Priority-based Q-value updates for efficient learning.
• Hybrid control integrating Rapidly-exploring Random Trees (RRT), Model Predictive Control (MPC), and Q-learning.

Main Results:

• Achieved a 92.4% success rate, 1.8% collision rate, and 6.7 cm parking alignment error in simulations.
• Demonstrated superior efficiency compared to DDPG, PPO, and DQN baselines.
• Confirmed significant contributions of individual modules through ablation studies (p < 0.01).

Conclusions:

• The proposed adaptive Q-learning framework is a viable proof-of-concept for dynamic parking scenarios.
• Light-weight tabular Q-learning with predictive extensions offers an efficient solution for autonomous parking.
• The method shows promise for real-world applications requiring precise maneuvering in complex environments.