Motion planning in complex urban environments: An industrial application on autonomous last‐mile delivery vehicles

Abstract

Abstract In this article, a motion planning framework for autonomous driving is explored to achieve unmanned last‐mile delivery vehicle application in complicated urban scenario. This approach can dramatically improve the intelligence, driving safety, driving robustness, and scalability of autonomous vehicles. In this framework, a specific High‐Definition (HD) map representation was proposed for last‐mile delivery applications, and a Route Planning layer generates a kinematically feasible reference line with smoothness condition based on Routing and HD Map components. Then a Scenario Planning layer considers routing results and both static and dynamic obstacles into account to select a corresponding scenario to execute, such as cruise on‐road scenario, cross intersection scenario, parking scenario, and so forth. Finally, a Trajectory Planning layer which is classified with trajectory generation and optimization modules is described. In the trajectory generation part, a rough path–speed profile and corresponding decisions, such as nudge, stop, yield, and so forth, are generated. Then, the rough path–speed profile is postprocessed by optimization algorithms in the trajectory optimization part. On the basis of the real road test results from thousands of times for JD.com's real autonomous delivery vehicle operations and approximate 115,475 km of fully autonomous driving mode in an urban scenario, the proposed motion planning framework demonstrates the efficiency of autonomous driving, improves the driving quality and reduces the manual intervention.