End-to-End Motion Planning Based on Visual Conditional Imitation Learning and Trajectory-Guide

End-to-end motion planning based on visual conditional imitation learning includes control prediction and trajectory planning. The former lacks global environment perception, the latter does not consider vehicle kinematic modeling, and both lack interpretability. This paper proposes a method that combines the two. Interpretability is provided for end-to-end networks by generating a semantic bird’s eye view. The high-level features of the bird’s eye view are fed into the trajectory planning branch and the control prediction branch, while the control prediction branch obtains trajectory guidance, making the two tightly integrated and mutually beneficial. In comparison experiments with vision-advanced methods LBC, TCP, and DLSSCIL on the NoCrash benchmark of the Carla simulation, the proposed method obtains an average gain of 13% in complex scenarios.