ECCV 2026: DriveTeach-VLA from Beihang and Tsinghua Improves Autonomous Driving VLA with Image Trajectories

A team from Beihang University, Tsinghua AIR, and Didi introduced DriveTeach-VLA, which uses driving-aware vision distillation and BEV-to-image trajectory mapping to improve the visual attention and trajectory prediction of autonomous driving VLA models. It achieves 90.4 PDMS on NAVSIM, and 92.7 with the Drivor selector, setting a new state of the art. The paper was accepted at ECCV 2026, with code and models open-sourced.