Automated Parking Path Planning Using Hybrid A* and Reinforcement Learning

Can a car find its own way into a parking bay? This project plans and drives automated parking manoeuvres in simulation and compares two classical path planners - Hybrid A* and RRT* on Reeds-Shepp curves, followed by a pure- pursuit controller - with reinforcement-learning agents (Soft Actor-Critic with Hindsight Experience Replay). On the project's own seeded perpendicular and parallel car parks Hybrid A* parks in 98.0 % and 90.5 % of 200 test scenarios each without touching anything, RRT* in 79.0 % and 61.0 % with contacts in 10.5 % and 7.5 %. In the highway-env parking-v0 simulator both planners park in every episode; the SAC agent trained on an empty lot reaches 98.0 % (93.5 % with parked cars), while the agent trained with parked cars failed (40.0 %). The planners win. A web application plans and drives any scenario live and checks the clearance. A simulation study and a teaching prototype, not a parking-assist function.

Category: Automobile Engineering

Automobile Engineering Project AU09 from ProTechSC Academy, with graphical abstract, video demonstration and project guidance.