Trajectory Optimization of Smart City Scenarios Using Learning Model Predictive Control

Abstract: Smart cities embrace cutting-edge technologies to improve transportation efficiency and safety. With the rollout of 5G and an ever-growing network of connected infrastructure sensors, real-time road condition awareness is becoming a reality. However, this progress brings new challenges. The communication and vast amounts of data generated by autonomous vehicles and the connected infrastructure must be navigated. Furthermore, different levels of autonomous driving (ranging from 0 to 5) are rolled out gradually and human-driven vehicles will continue to share the roads with autonomous vehicles for some time. In this work, we apply a data-driven control scheme called Learning Model Predictive Control (LMPC) to three different smart city scenarios of increasing complexity. Given a successful execution of a scenario, LMPC uses the trajectory data from previous executions to improve the performance of subsequent executions while guaranteeing safety and recursive feasibility. Furthermore, the performance from one execution to another is guaranteed to be non-decreasing. For our three smart-city scenarios, we apply a minimum time objective and start with a single vehicle in a two-lane intersection. Then, we add an obstacle on the lane of the ego vehicle, and lastly, we add oncoming traffic. We find that LMPC gives us improved traffic efficiency with shorter travel. However, we find that LMPC may not be suitable for real-time training in smart city scenarios. Thus, we conclude that this approach is suitable for simulator-driven, offline, training on any trajectory data that might be generated from autonomous vehicles and the infrastructure sensors in future smart cities. Over time, this can be used to construct large data sets of optimal trajectories which are available for the connected vehicles in most urban scenarios.