Learning MPC for Interaction-Aware Autonomous Driving: A Game-Theoretic Approach

Abstract

We present a novel control strategy for controlling autonomous vehicles in general traffic situations which accounts for the mutual interactions between the controlled vehicle and other road users. More specifically, the interaction is modelled as a generalized potential game, where each road user is assumed to minimize a shared cost function subject to shared (collision avoidance) constraints. The shared cost allows the controlled vehicle to cooperate with other road users, while safety guarantees follow from the imposed hard collision avoidance constraints and the introduction of a model predictive control feedback scheme. In the case where the incentives and constraints of other road users, i.e., human drivers, are unknown, we propose a natural and practical methodology for learning this information online from observed data and incorporating it directly into the solution methodology for the game formulation. Extensive numerical simulations in a realistic highway merging scenario have been performed, verifying the practical usability of the developed methodologies.

Safe, learning-based MPC for highway driving under lane-change uncertainty: A distributionally robust approach

“…Similarly to [45], we conclude that collision occurs whenever at least one corner of at least one vehicle is at distance zero from the other vehicle. For each target vehicle in the environment, this leads to 8 constraints similar to (5) with functions h of the form (9); one for every corner of every vehicle.…”