Predictive Neural Motion Planner for Autonomous Driving Using Graph Networks

Recent advances in digital twin and parallel intelligence (DTPI) enable the mapping of the physical world to a high-fidelity virtual representation and facilitate intelligent prediction and decision-making for autonomous vehicles and intelligent transportation systems. In the context of DTPI, in this study, we investigate trajectory-prediction-enabled motion planning for autonomous vehicles using deep neural networks. We first implement a motion planner using a neural network as an approximation of traditional planners. The inputs to the baseline planner include the current states of the ego and its surrounding agents and a shared map. The planner produces a five-second trajectory for the ego vehicle considering the current situation. Subsequently, we generalize the baseline to consider the historical states of the ego and its surrounding agents. Using the generalized planner, we investigate the impacts of the history horizon on planning performance. We next investigate how the future motions of the surrounding agents of the ego affect the planner and observe improvement in planning. This demonstrates that knowledge of the future trajectories of other agents is useful for planning. However, we do not have access to ground-truth future motions for inference. Finally, we investigate how the future can be approximated through prediction and how the prediction quality affects planning performance.