Towards Safe and Efficient Modular Path Planning using Twin Delayed DDPG

Path planning is an essential function in autonomous vehicles. Several Reinforcement Learning (RL) algorithms were explored for achieving the path planning task. Challenges include generating optimal vehicle paths while maintaining high safety standards and low training time. In this paper, we present a Twin Delayed Deep Deterministic Policy Gradient (TD3) RL algorithm as a modular path planner. This path planner utilizes an occupancy grid map training environment, integrated with the Intel Responsibility Sensitive Safety (RSS) rules for planning high safety standards paths. The resulting TD3 algorithm is compared to its predecessor, the Deep Deterministic Policy Gradient (DDPG) RL algorithm, in terms of the training time. The results show that our proposed path planner successfully reduced the needed training time by 31% in comparison to the DDPG-based path planner. Additionally, CARLA simulations were provided for validation, showing the high safety standards in the planned paths.