Learning from Demonstration: Situation-Adaptive Lane Change Trajectory Planning for Automated Highway Driving

Trajectory planning is one of the key challenges to enable automated driving system on the road. In structured highway environment, lane change trajectory planning algorithm for automated driving has generally been formulated as optimizing complex cost functions which should be manually specified and tuned by experts. However, considering the highly dynamic environment in highway automation, a wide variety of cost functions may be designed to work in different driving situations, making the hand-tuning a recurring process and requires rich experience. This research proposes the driving situation-adaptive lane change trajectory planning approach by learning from expert demonstrations. Specifically, in offline learning process, the trajectory planning cost functions database is constructed by correlating the driving situation assessment result with the corresponding situation-specific planning cost function which can be learned from expert demonstration trajectory automatically via feature-based inverse reinforcement learning technique. In online trajectory optimization process, the lane change environment is evaluated by Bayesian Network firstly and then the situation-dependent cost function can be synthesized from database. The final planned lane change trajectory can be effectively calculated by optimizing the synthesized cost function. Simulation experiments in different driving environments demonstrate that the proposed method is capable of learning distinct cost functions and planning driving situation-adaptive lane change trajectory using data from expert demonstrated trajectories.