DriveStack-VLA: Render-Teacher Alignment for BEV-Based DeepStack Vision-Language-Action Model

2026-06-22

AI Summary

Existing vision-language-action driving models inadequately capture metric geometry, bird’s-eye-view (BEV) scene structure, and critical safety cues. This work proposes a driving policy framework grounded in large vision-language models, featuring a Render-Teacher Alignment mechanism that aligns perceptual attention between real images and rasterized renderings. It further enhances spatial understanding through DeepStack-style BEV fusion and incorporates a head self-critique module to refine trajectory selection and fine-tuning. The approach substantially improves spatial reasoning and safety of driving policies, achieving 91.6 PDMS and 91.0 EPDMS (with human penalty filtering) on NAVSIMv1 and v2, respectively, and attaining a driving score of 79.49 with a 56.36% success rate on the Bench2Drive closed-loop benchmark.

Abstract

Vision-Language-Action driving models convert a pretrained Vision-Language Model into a driving policy, allowing them to use world knowledge and follow language guidances. However, existing VLA driving models still lack driving-oriented spatial intelligence: their policies are mainly grounded on perspective image tokens and language priors, while precise motion planning requires metric geometry, top-down scene structure, and attention to safety-critical perceptual cues. This limitation makes current models vulnerable to weak visual geometry modeling and perceptual coverage in expert demonstrations. In this paper, we present DriveStack-VLA, a framework built upon a large VLM backbone. To strengthen the spatial grounding of VLA driving, we develop dual visual modeling components. We inject a Bird-Eye-View representation into the Large Language Model decoder through a DeepStack-style connection, and propose Render-Teacher Alignment to align the perceptual focus of real images with that of rasterized images. Furthermore, to bridge the gap in multimodal trajectory selection, we introduce a head-based self-critique module that ranks sampled trajectories and conditionally refines the best one. DriveStack-VLA achieves 91.6 PDMS on NAVSIMv1, 91.0 EPDMS on NAVSIMv2 (with the human penalty filter enabled), and a driving score of 79.49 with a success rate of 56.36\% on the closed-loop Bench2Drive. More visualizations are available on our project page: https://anonymous.4open.science/w/drivestack-vla/.

Problem

• Vision-Language-Action
• spatial intelligence
• Bird-Eye-View
• motion planning
• perceptual grounding

Innovation

• Bird-Eye-View (BEV)
• Render-Teacher Alignment
• DeepStack
• Vision-Language-Action (VLA)
• self-critique trajectory ranking