Model-Based Predictive Control and Reinforcement Learning for Planning Vehicle-Parking Trajectories for Vertical...

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Automatic parking systems often face challenges with long trajectory planning times and convergence difficulties.
• Existing methods may require extensive training and struggle with efficient path generation.

Purpose of the Study:

• To develop an efficient trajectory planning method for automatic parking.
• To address issues of long planning times and slow convergence in autonomous parking.

Main Methods:

• A two-stage approach combining Model Predictive Control (MPC) for initial trajectory tracking and Proximal Policy Optimization (PPO) for reinforcement learning-based parking.
• Utilizing a four-dimensional reward function to guide neural network parameter adjustment and minimize invalid actions.

Main Results:

• The MPC controller demonstrated accurate trajectory tracking with smooth, continuous vehicle movement.
• The PPO-based method achieved significantly shorter learning times (30-37.5% of DDPG/TD3) and faster convergence (75% and 68% fewer iterations than DDPG/TD3).

Conclusions:

• The proposed method effectively overcomes slow convergence and long training times in parking trajectory planning.
• This approach enhances the timeliness and efficiency of autonomous parking operations.