Learning Adaptive Parking Maneuvers for Self-driving Cars

This paper addresses the autonomous parking for a vehicle in environments with static and dynamic obstacles. Although parking maneuvering has reached the level of fully automated valet parking, there are still many challenges to realize the parking motion planning in the presence of dynamic obstacles. One of the most famous autonomous driving platforms is the Baidu Apollo platform. In the Apollo platform, this problem is solved using the classic method hybrid A*. However, this method has two main downsides. Firstly, it generates in some parking scenarios, trajectories that consist of many partitions with different gear types and sizes. Such trajectories are intractable by a self-driving car when testing the Apollo planner on more realistic data coming from a simulator such as SVL. Secondly, the built-in algorithm does not have the ability to interact with dynamic obstacles, which might lead to a collision in some critical parking scenarios. To overcome these issues, we proposed a method based on reinforcement learning, which uses the RL-policy (from POLAMP) allowing us to take into account the kinematic constraints of the vehicle, static and dynamic obstacles. The proposed method was fully integrated into the Apollo platform with developed Cyber RT nodes, which were used for publishing the parking trajectory from our algorithm to the SVL simulator through a ROS/Cyber bridge. The final model demonstrates transferability to the previously unseen experimental environments and flexibility with respect to built-in hybrid A*.