Imitation Learning of Hierarchical Driving Model: From Continuous Intention to Continuous Trajectory

One of the challenges to reduce gap between machine and human level driving is how endow system with learning capacity deal coupled complexity environments, intentions, dynamics. In this letter, we propose a hierarchical model explicit models continuous intention dynamics, which decouples in observation-to-action reasoning data. Specifically, module takes perception generate potential map encoded obstacles intentions. Then, regarded as condition, together current trajectory output by function approximator network, whose derivatives can be used for supervision without additional parameters. Finally, our method validated both datasets stimulation, demonstrating that has higher prediction accuracy displacement velocity generates smoother trajectories. Our also deployed on real vehicle loop latency, validating its effectiveness. To best knowledge, first work produce using network. code available at https://github.com/ZJU-Robotics-Lab/CICT.