ECCV 2026｜自驾VLA Scaling有戏了，北航清华DriveTeach-VLA：用图像轨迹打通驾驶场景与基模预训练
Ranking
Observed public metrics from 1 member.
Merged summary
TL;DR - DriveTeach-VLA (Beihang, Tsinghua AIR, DiDi; ECCV 2026) is an autoregressive vision-language-action model for autonomous driving that fixes VLAs' misplaced visual attention and bridges BEV trajectories to a multimodal base model's native image-reading ability, reaching 90.4 PDMS on NAVSIM. It matters because it suggests VLA scaling can come from better spatial supervision rather than more chain-of-thought text or bigger models.
- DVD (Driving-aware Vision Distillation): GroundingDINO labels vehicles/pedestrians/barriers/traffic lights; a teacher ViT sees bbox-annotated images while a student ViT sees raw images, with Swin-style block-wise pooled feature distillation transferring the attention correction. Worth ~1.4 PDMS, and robust to 20% bbox noise (random drop, jitter).
- 2D-TGP (2D Trajectory-Guided Prompt): expert BEV trajectories are projected via camera intrinsics/extrinsics into image pixel coordinates and fed back as text prompts, making trajectory geometry legible to the pretrained base model. Two Qwen2.5-VL-3B models split the work — a TGP-Prompter predicts the 2D trajectory, a TGP-Planner consumes it to output the BEV trajectory.
- Ablations: Qwen2.5-VL-3B baseline 84.8 PDMS → 86.4 with VQA+CoT → 87.3 with DVD → 88.2 with 2D-TGP → 90.4 PDMS / 85.4 EPDMS after GRPO behavior alignment; nuScenes L2 error 0.3. Planner stays at 3B, so gains come from supervision and the intermediate interface, not scale.
- Sampling headroom: with no diffusion or continuous MLP action head, the model keeps LLM sampling properties — 12 candidate trajectories reranked by the Drivor selector push results to 92.7 PDMS / 89.0 EPDMS (not deployable as-is, but an upper bound for distillation). Training uses teacher-forced ground-truth 2D-TGP while inference uses predicted prompts; the authors verify performance degrades gracefully under that gap.
Sources (1)
ECCV 2026｜自驾VLA Scaling有戏了，北航清华DriveTeach-VLA：用图像轨迹打通驾驶场景与基模预训练
TL;DR - DriveTeach-VLA (Beihang, Tsinghua AIR, DiDi; ECCV 2026) is an autoregressive vision-language-action model for autonomous driving that fixes VLAs' misplaced visual attention and bridges BEV trajectories to a multimodal base model's native image-reading ability, reaching 90.4 PDMS on NAVSIM. It matters because it suggests VLA scaling can come from better spatial supervision rather than more chain-of-thought text or bigger models.
- DVD (Driving-aware Vision Distillation): GroundingDINO labels vehicles/pedestrians/barriers/traffic lights; a teacher ViT sees bbox-annotated images while a student ViT sees raw images, with Swin-style block-wise pooled feature distillation transferring the attention correction. Worth ~1.4 PDMS, and robust to 20% bbox noise (random drop, jitter).
- 2D-TGP (2D Trajectory-Guided Prompt): expert BEV trajectories are projected via camera intrinsics/extrinsics into image pixel coordinates and fed back as text prompts, making trajectory geometry legible to the pretrained base model. Two Qwen2.5-VL-3B models split the work — a TGP-Prompter predicts the 2D trajectory, a TGP-Planner consumes it to output the BEV trajectory.
- Ablations: Qwen2.5-VL-3B baseline 84.8 PDMS → 86.4 with VQA+CoT → 87.3 with DVD → 88.2 with 2D-TGP → 90.4 PDMS / 85.4 EPDMS after GRPO behavior alignment; nuScenes L2 error 0.3. Planner stays at 3B, so gains come from supervision and the intermediate interface, not scale.
- Sampling headroom: with no diffusion or continuous MLP action head, the model keeps LLM sampling properties — 12 candidate trajectories reranked by the Drivor selector push results to 92.7 PDMS / 89.0 EPDMS (not deployable as-is, but an upper bound for distillation). Training uses teacher-forced ground-truth 2D-TGP while inference uses predicted prompts; the authors verify performance degrades gracefully under that gap.