Personalized Lane Change Planning and Control by Imitation Learning from Drivers

In this article, we propose a novel personalized planning and control approach for lane change assistance system which can efficiently learn a model prediction control (MPC)-based driver-specific lane-changing policy via end-to-end imitation learning from a few driver demonstrations. Specifically, we build a novel learnable predictive model of the vehicle-driver system and design an adaptable cost function for the MPC-based lane change controller. We then calculate the gradient of the imitation loss with respect to the personalization parameters of the model and cost function via differentiating the optimality conditions, and update those parameters to minimize the imitation loss in an end-to-end fashion. A semi-physical simulation on a driving simulator and a closed-loop test on a real vehicle are conducted to validate the learning ability and personalized control performance. The results show that 1) the proposed method can automatically implement both the generalized and the personalized lane change planning and control by learning from demonstration data; 2) the proposed controller can adapt to different driver-specific behaviors; and 3) the proposed approach outperforms the model-free learning approach in terms of imitation accuracy, interpretability, data efficiency, and generalized performance.