Safe and Human-Like Trajectory Planning of Self-Driving Cars: A Constraint Imitative Method

Safe and human-like trajectory planning is crucial for self-driving cars. While model-based planning has demonstrated reliability, it is beneficial to incorporate human demonstrations and align the results with human behaviors. This work aims at bridging the gap between model-based planning and driver imitation by proposing a constraint imitative trajectory planning method (CITP). CITP integrates artificial potential field and dynamic movement primitives, which have achieved both the ability to imitate human demonstrations as well as ensure safety constraints. During the planning process, CITP first encodes human demonstrations, local driving target, and traffic obstacles as attractive or repulsive effects, and then the trajectory planning problem is solved through model predictive optimization. To address the dynamics of traffic scenarios, a hierarchical planning strategy is proposed based on the division of planning process. CITP is designed with five modules, including LSTM-based target generation, encoding attractive and repulsive effects with target, demonstrations and obstacles, and trajectory planning with model predictive optimization. Data collection and experiments are carried out based on CARLA driving simulator, and the effectiveness in terms of both safety and consistency with human behavior are reported.