Learning constraint identification and classification for online model predictive planning at unsignalized intersections for autonomous vehicles

Abstract

Decision making and planning through traffic in an unsignalized intersection is one of the most challenging tasks for a self-driving vehicle. Rule-based planning systems have been tested in real world implementations but this method relies on hand-engineering the rules which can suffer from being too conservative and unable to adapt to complex scenarios outside of it’s designed rule. On the other hand, Model Predictive Planning (MPP) is a powerful optimization based approach that is robust and is able to find the optimal trajectory within its constraints. However, implementing MPP for autonomous driving requires solving an online optimization problem that is generally non-convex, which significantly increase the computation burden for any optimization solver and limits the use of MPP for real time applications. In this work, MPP is used for longitudinal speed planning in an unsignalized intersection and a deep learning based framework which can reduce the computation burden for the MPP is proposed. Since the heavy computational load in model predictive-based approaches comes from the non-convex safety constraint with the surrounding vehicles in the environment, the proposed framework can reduce the complexity by finding a convex set from the non-convex constraints. This is done by 1.) Identifying the active constraints for the optimization problem (selecting only the target vehicles that are critical to safety) and 2.) Classifying whether a constraint is an upper bound or lower bound constraint (deciding whether the optimal solution is to go before or after a vehicle in the intersection). A deep neural network (DNN) will be part of a high level decision scheme that is used to identify and classify these constraints and will be combined with the low level MPP. To identify the correct convex set, two DNN training schemes that incorporates feedback from the MPP is explored. Finally, the approach is evaluated with several baselines in an unsignalized intersection scenario.