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

A new framework called DriveStack-VLA aims to improve the spatial reasoning of Vision-Language-Action driving models by incorporating a bird’s-eye-view representation and a self-critique module, according to a preprint posted on arXiv [1]. The paper, submitted on 23 June 2026, details how existing VLA driving models convert a pretrained Vision-Language Model into a driving policy, enabling them to use world knowledge and follow language instructions [1]. The authors argue that current models lack driving-oriented spatial intelligence because their policies rely on perspective image tokens and language priors, whereas precise motion planning requires metric geometry, top-down scene structure, and attention to safety-critical perceptual cues [1]. This limitation makes the models vulnerable to weak visual geometry modeling and perceptual coverage in expert demonstrations [1]. To address this, DriveStack-VLA is built upon a large VLM backbone and introduces dual visual modeling components [1]. A Bird-Eye-View representation is injected into the Large Language Model decoder through a DeepStack-style connection, and a technique called Render-Teacher Alignment aligns the perceptual focus of real images with that of rasterized images [1]. The framework also includes a head-based self-critique module that ranks sampled trajectories and conditionally refines the best one, bridging the gap in multimodal trajectory selection [1]. On the NAVSIMv1 benchmark, DriveStack-VLA achieves a PDMS score of 91.6, and on NAVSIMv2 it records an EPDMS of 91.0 with the human penalty filter enabled [1]. In closed-loop testing on the Bench2Drive dataset, the model posts a driving score of 79.49 and a success rate of 56.36% [1]. The preprint appears on arXiv, an open-access repository of electronic preprints that is not peer-reviewed and which, as of November 2024, receives about 24,000 submissions per month [6]. The work is categorized under Computer Vision and Pattern Recognition [1].