DriveTeach-VLA: Tsinghua & Beihang Model Hits SOTA by Teaching AVs to Look
机器之心 · wechat · 2026-08-10
A joint team from Tsinghua University, Beihang University, and Didi has proposed the DriveTeach-VLA model to address the issue of existing autonomous driving Vision-Language-Action (VLA) models over-relying on text reasoning while ignoring visual elements. The paper has been accepted by ECCV 2026.
The researchers point out that while LLMs can describe driving scenarios well, what truly determines safety is their visual attention when generating trajectories. To fix this, they designed three modules: Driving-aware Vision Distillation (DVD) to focus the model on key traffic objects; 2D Trajectory-Guided Prompts (2D-TGP) to map abstract Bird's-Eye-View (BEV) trajectories into pixel space to activate the base model's inherent image-reading abilities; and finally, behavior alignment. The model achieves 90.4 PDMS on the NAVSIM benchmark, and reaches 92.7 PDMS with a trajectory selector, setting a new State-of-the-Art (SOTA).
More from Research
- Leaked Architecture of ~400B MoE Model with Aggressive GQA Sparks Interest — teortaxesTex · 2026-08-10
- Research: Unconditional Prediction Accuracy Isn't Always the Right Objective in AI Decision Processes — joshgans · 2026-08-10
- Synthetic Query Probing: Comparing Similarity Spaces Across Embeddings — pppeer · 2026-08-10
- Gödel, Escher, Bach Predicted AI: Gradient Descent as the Ultimate 'Strange Loop' — AymericRoucher · 2026-08-10
- OneEmo: A Unified Multimodal Reasoning Model for Emotion Perception — Jiahao Huang · 2026-08-10
- Making Knowledge Distillation Cheap Enough to Run at Scale — Hugging Face Blog · 2026-08-10