| Providing safe,efficient and comfortable motion planning strategy remains challenges for autonomous vehicles.Most current mainstream methods such as model predictive control(MPC)often fail in complex interactive environments due to the fact that they do not account for the interactions between the ego and nearby traffic and that the time complexity of such algorithms is high for on-board vehicle computing devices.As an alternative,deep reinforcement learning(DRL)can learn how to best interact with the surrounding environment without an explicit model for nearby drivers’ behaviors and is able to provide optimal control under real-time considering almost all computation can be carried out offline.Nevertheless,the lack of explainability for DRL-based solutions may prevent their large-scale application in industrial autonomous vehicle tasks.Furthermore,the DRL method tends to be unsafe and brittle to scenarios not encountered in training.In this context,a novel DRL-based motion planning method is proposed that explicitly takes explainability and robustness into account.Specifically,unlike traditional DRL-based solutions which executes from perception to control in an end-to-end manner,this work decouples the layer of motion planning from the end-to-end solution by adding uncertainty-aware interval prediction to compute the set of states that can be reached over planning time horizon.On this basis,a robust control framework that aims to guarantee safety with the worst-case performance of the system is detailed.To validate the proposed algorithm,the task of an autonomous vehicle merging on to a highway from an on-ramp is simulated in SUMO.The results show that the proposed motion planning method combines the advantages of optimization-based and DRL methods and achieves balanced performance in smoothness,computational efficiency,explainability and robustness.The main research contents of this paper are as follows:1)In contrast to traditional DRL-based solutions which are often in a form of artificial neural network and thus are difficult to know how all the individual neurons work together to arrive at the final output,this work decouples the layer of motion planning from the end-to-end solution by adding a hybrid uncertainty-aware prediction model.By rolling out the potential set of states that can be reached over the predicted time horizon into the S-T space,the trajectories of both nearby traffic and ego vehicle can be generated and visualized.This S-T map can help evaluate whether the trajectory of the motion planning is safe in a long run,thereby increasing the explainability of the DRL-based method.2)Considering the fact that DRL-based methods only solve the autonomous vehicle control problems by approximating the expectation of samples and the notorious simulation-to-real gap,there is no strict safety guarantee of the DRL-based policy in both theory and practice.To ensure the safety performance,building on the aforementioned planned DRL while considering vehicle dynamics constraints,a robust control framework which aims to guarantee safety with the worst-case performance of the nearby traffic system is proposed. |
