SpanVLA: Efficient Action Bridging and Learning from Negative-Recovery for Autonomous Driving

Forum topic · 小凯 · 2026-04-23

Summary

SpanVLA is an end-to-end autonomous driving framework that integrates autoregressive vision-language reasoning with a flow-matching action expert. To address the high latency and limited robustness of existing Vision-Language-Action (VLA) models that generate actions autoregressively, SpanVLA introduces an efficient bridge that leverages VLM vision and reasoning guidance to plan future trajectories via a flow-matching policy conditioned on historical trajectory initialization, significantly reducing inference time. The authors further propose a GRPO-based post-training method that teaches the model to learn from positive samples while also avoiding typical negative behaviors and learning recovery behaviors. A new dataset, mReasoning, is introduced, focusing on complex reasoning-heavy driving scenarios with negative-recovery samples. Extensive experiments on NAVSIM v1 and v2 demonstrate competitive planning performance and robustness across diverse scenarios. Paper on arXiv: 2604.19710.

Paper Overview

Field: Computer Vision / Autonomous Driving Authors: Zewei Zhou, Ruining Yang, Xuewei Qi, Yiluan Guo, Sherry X. Chen, Tao Feng, Kateryna Pistunova, Yishan Shen, Lili Su, Jiaqi Ma arXiv: 2604.19710

Abstract

Vision-Language-Action (VLA) models offer a promising autonomous driving paradigm for leveraging world knowledge and reasoning capabilities, especially in long-tail scenarios. However, existing VLA models often struggle with the high latency in action generation using an autoregressive generation framework and exhibit limited robustness.

This paper proposes SpanVLA, a novel end-to-end autonomous driving framework integrating autoregressive reasoning with a flow-matching action expert:

Efficient action bridging: SpanVLA introduces an efficient bridge to leverage the vision and reasoning guidance of the VLM, planning future trajectories with a flow-matching policy conditioned on historical trajectory initialization, which significantly reduces inference time.
Learning from negative-recovery: A GRPO-based post-training method improves performance and robustness, enabling the VLA model to learn not only from positive samples but also how to avoid typical negative behaviors and learn recovery behaviors.
mReasoning dataset: A new real-world driving reasoning dataset focused on complex, reasoning-intensive scenarios with negative-recovery samples.

Extensive experiments on NAVSIM (v1 and v2) demonstrate the competitiveness of SpanVLA, and qualitative results across diverse scenarios highlight the model's planning performance and robustness.

SpanVLA: Efficient Action Bridging and Learning from Negative-Recovery for Autonomous Driving

Summary

Paper Overview

Abstract

Links

Tags