Reinforcement Learning for Autonomous Parking

Author(s) Information

Ravi Panchal, California Polytechnic State University, San Luis ObispoFollow

College - Author 1

College of Engineering

Department - Author 1

Electrical Engineering Department

Advisor

Ria Kanjilal, College of Engineering, Computer Engineering Department

Date

10-2025

Abstract/Summary

This research project proposes the development of an adaptive reinforcement learning (RL)-based parking system designed to handle complex parking maneuvers that are often unaddressed in existing research. Unlike previous works that focus on isolated maneuvers such as reverse parking or parallel parking, this work presents a flexible framework where multiple parking types (reverse, parallel, diagonal) are addressed through independent agents and then integrated into a cohesive system. The primary gap addressed by this work is the integration of diverse parking maneuvers into a unified RL framework with adaptability to dynamic and varied parking environments. Additionally, this project introduces curriculum learning to enhance training efficiency and domain randomization to improve robustness against environmental variations. This multi-maneuver capability allows for a broader, more scalable application in smart parking systems that can handle different vehicle types and parking scenarios without requiring retraining from scratch. Building on the foundational work of RL-based reverse parking, this project aims to generalize the approach to multiple parking tasks while enhancing robustness through novel training methodologies. The proposed system is trained using the Proximal Policy Optimization (PPO) algorithm within customized simulation environments using OpenAI Gymnasium and HighwayEnv frameworks. A key contribution of this research is the development of an adaptive control architecture that dynamically selects the appropriate maneuver type based on the parking scenario. The project will culminate in a comprehensive simulation-based evaluation demonstrating the system’s ability to effectively handle varied parking tasks with high success rates. The results will inform future extensions to real-world scenarios and contribute valuable insights for integrating diverse control tasks within a single RL framework.

October 1, 2025.