Autonomous Valet Parking with Asynchronous Advantage Actor-Critic Proximal Policy Optimization

Autonomous valet parking (AVP) is one of the main features of autonomous vehicles. Autonomous parking system is essential to reduce waiting time to search for available lot then parking, particularly in the smart city. One of the drawbacks is the inevitable control errors due to the vehicle non-linear dynamics. Hence, this brings the motivation to implement reinforcement learning into autonomous valet parking vehicles. A Proximal policy optimization with asynchronous advantage actor-critic (A3C-PPO) is proposed for autonomous valet parking. This system is simpler and can solve problems that has continuous state-action space which includes autonomous valet parking. For the agent to yield better final rewards, manual hyperparameter tuning is conducted which are mini-batch size tuning and generalized advantage estimate (GAE) factor tuning. A3C enable parallel computing and speed up the agent training.