Tuning an MPC-based Motion Planner using Imitation Learning

Abstract: Model Predictive Control, MPC, is one of the most commonly used controllers today. They are well understood with a vastly researched theoretical background. One of the core strengths of MPC is that it is possible to provide safety guarantees for complex systems through the constraints enforced in the optimization problem. This makes them useful for autonomous vehicle trajectory planning. However, as system complexity goes up, so does the number of states and thus also the number of weight parameters. Tuning these parameters is a difficult, tiresome and time-consuming task that could be automated from human driving data. The aim of this thesis is to tune the cost function of an MPC-based motion planner for in- lane driving from human demonstrations. The motion planner used is a non-linear path-following MPC designed with a cost function of 6 features commonly used for autonomous in-lane driving. The cost function weights are learned from real human data through the framework of Maximum Margin Planning, which is an inverse reinforcement learning algorithm. The proposed framework is evaluated and it is shown that it can successfully learn the underlying behavior of drivers in selected data sets. However, it is also shown that the non-convexity of the motion planner becomes problematic for other data sets. Therefore, a simpler convex MPC motion planner is designed, for which the proposed framework is shown to successfully tune weight parameters based on expert driving data.