Paper Overview
- Field: Computer Vision (CV)
- Authors: Zewei Zhou, Ruining Yang, Xuewei Qi, Yiluan Guo, Sherry X. Chen, Tao Feng, Kateryna Pistunova, Yishan Shen, Lili Su, Jiaqi Ma
- Published: 2026-04-21
- 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 that integrates autoregressive reasoning with a flow-matching action expert:
1. Efficient action bridging: SpanVLA introduces an efficient bridge to leverage the vision and reasoning guidance of the VLM, planning future trajectories via a flow-matching policy conditioned on historical trajectory initialization, which significantly reduces inference time.
2. Learning from negatives and recovery: A GRPO-based post-training method is proposed to further improve 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.
3. mReasoning dataset: The authors introduce mReasoning, a new real-world driving reasoning dataset focused on complex, reasoning-intensive scenarios and negative-recovery samples.
Extensive experiments on NAVSIM (v1 and v2) demonstrate the competitiveness of SpanVLA. Qualitative results across diverse scenarios further highlight the model's planning performance and robustness.
---
*Auto-collected on 2026-04-23.*