Scenario-driven Motion Planning Model Selection in Autonomous Driving Systems

Overview

As an essential module in autonomous driving systems, motion planning is responsible for generating a collision-free, kinematic feasible, and comfortable driving trajectory given the inputs from environment perception, prediction, and global routing. With safety as the top concern, motion planning models need to generate a feasible trajectory in a timely manner and be responsive to changing traffic scenarios.

Many motion planning models have been developed in recent years with the aim to provide the best average performance under all possible scenarios where performance is usually evaluated from the ego vehicle's perspective, referred to as microscopic performance. The current research development in motion planning cannot address the following challenges in adopting motion planning models for real autonomous driving systems. Firstly, due to the complexity of traffic scenarios, it is impossible to develop a one-size-fits-all model for all traffic scenarios. On the other hand, it is not necessary to adopt only one motion planning model either. The challenge from this perspective is how to select the most appropriate motion planning models according to traffic scenarios. Secondly, when evaluating motion planning models, we should consider not only the performance of the ego vehicle, but also its impact on traffic flow, referred to as macroscopic performance. The challenge from this perspective is how to evaluate the impact of a motion planning model on its current traffic scenario, predict its impact on the following traffic scenarios, and proactively incorporate such impact into global routing.

To address the aforementioned challenges, this project will by no means develop another motion planning model, but consolidate the existing efforts in motion planning to select the right motion planning models for different scenarios. To this end, we will first develop a framework to comprehensively evaluate the microscopic and macroscopic performance of existing models and recommend the most appropriate model for any given scenario. To capture the impact of motion planning model selection on traffic scenarios in the following lane segments, we will develop a traffic scenario prediction model to predict the traffic flow in each following lane segment, which will be used as the input of existing routing algorithms so that microscopic and macroscopic performances can be better aligned. To ensure the responsiveness of motion planning, we will develop an online motion planning model selection framework to handle the cases where the pre-selected motion planning model cannot generate a feasible trajectory in a timely manner due to changing traffic scenarios.