Data Efficient Reinforcement Learning for Integrated Lateral Planning and Control in Automated Parking System.

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Reinforcement learning (RL) shows promise for automated parking systems (APSs) by integrating planning and control.
• Existing model-free RL methods require extensive interaction, while model-based RL struggles with continuous learning in APS.

Purpose of the Study:

• To develop a data-efficient, model-based reinforcement learning method for enhanced automated parking.
• To improve the learning speed and performance of RL algorithms in APS applications.

Main Methods:

• A truncated Monte Carlo tree search evaluates parking states and selects moves.
• Two artificial neural networks are trained for move probability and state reward using self-generated data.
• Data efficiency is boosted through weighted exploration, adaptive schemes, and experience augmentation with imaginary rollouts.

Main Results:

• The integrated RL method effectively coordinates longitudinal and lateral motion for parking in confined spaces.
• Simulations demonstrate adaptability to vehicle model changes and convergence within few iterations.
• Real-world experiments confirm superior parking attitude and success rates compared to simulation-based reward generation.

Conclusions:

• The proposed data-efficient RL method significantly improves APS performance and learning speed.
• The approach offers enhanced adaptability and effectiveness for real-world autonomous parking scenarios.
• This research advances the integration of advanced AI techniques in intelligent vehicle systems.