Deep Reinforcement Learning combined with RRT for trajectory tracking of autonomous vehicles.

Abstract

Sample inefficiency is a long-standing problem in Deep Reinforcement Learning based algorithms, which shadows the potential of these techniques. So far, the primary approach for tackling this issue is prioritizing the gathered experiences. However, the strategy behind collecting the experiences received less attention, but it is also a legitimate approach for prioritizing. In this paper, the Rapidly exploring Random Trees algorithm and Deep Reinforcement Learning are combined for the trajectory tracking of autonomous vehicles to mitigate the issues regarding sample efficiency. The core of the concept is to utilize the tremendous explorational power of RRT for covering the state space via experiences for the Agent to diversify its training data buffer. The results demonstrate that this approach outperforms the classic trial-and-error-based concept according to several performance indicators.