Reasoning-VLA: An Efficient and Spatial-Guided General Vision-Language-Action Reasoning Model for Autonomous Driving

Vision-Language-Action (VLA) models have recently shown strong decision-making capabilities in autonomous driving. However, existing VLAs often struggle with achieving efficient inference and generalizing to novel autonomous vehicle configurations and driving scenarios. In this paper, we propose Reasoning-VLA, a general and efficient action-generation VLA framework. The proposed model employs a set of learnable action queries, implicitly guided by predefined spatial representations to enhance spatial awareness. These learnable queries interact with reasoning-enhanced vision–language features to generate continuous action trajectories in parallel. To promote robust generalization, we consolidate eight publicly available autonomous driving datasets into a standardized, Chain-of-Thought reasoning–based, and easy-to-use data format for model training. Leveraging both supervised learning and reinforcement learning fine-tuning, extensive empirical evaluations across multiple benchmarks demonstrate that Reasoning-VLA achieves state-of-the-art performance, strong generalization capability, and the excellent inference speed with parallel decode.

Recent VLA in AD have two big problems: they often work slowly when making decisions, and they struggle to adapt to new types of self-driving cars or new driving situations (like different roads or weather). In our paper, we created a new VLA framework called Reasoning-VLA to fix these issues. Our model uses special, adjustable “action guides” that are helped by a clear sense of space (like knowing where other cars or roads are) to make better decisions. These guides work with improved vision and language understanding to plan driving actions (like turning or speeding up) all at once, which makes the model faster. To make sure Reasoning-VLA can adapt well to new situations, we combined eight different public datasets of self-driving data into a simple, easy-to-use format that teaches the model to think step-by-step. We trained the model using two common methods: one where it learns from example data, and another where it learns by trying actions and getting feedback. Tests on multiple benchmarks show that Reasoning-VLA works better than other models, adapts well to new scenarios, and makes decisions quickly by planning actions in parallel.