DriveTeach-VLA: Teaching Vision-Language-Action Models What to See and Where to Look

jiqizhixin · x · 2026-08-18

Researchers from Beihang University, Tsinghua University, and DiDi present DriveTeach-VLA, a new training framework for autonomous driving models. Current AI drivers get overwhelmed by text-heavy reasoning data, losing focus on spatial clues needed for safe navigation.

DriveTeach-VLA changes this by teaching models a two-step vision: first, identifying what visual details matter for driving, then guiding attention based on real, feasible driving paths. This links seeing with acting. The result is SOTA performance on NAVSIM and nuScenes benchmarks, with code available on GitHub.

Original post →

More from Embodied

Embodied channel →