AutoDrive-R$^2$: Incentivizing Reasoning and Self-Reflection Capacity for VLA Model in Autonomous Driving

TL;DR

AutoDrive-R² introduces a two-stage training framework for Vision-Language-Action autonomous driving that combines structured chain-of-thought reasoning with self-reflection and a physics-grounded GRPO RL regime to ensure feasible trajectory planning. A novel nuScenesR²-6K CoT dataset with self-validation guides foundational reasoning, while physics-informed rewards align planning with spatial, dynamic, and temporal constraints. Empirical results on nuScenes and Waymo demonstrate state-of-the-art trajectory accuracy and strong zero-shot generalization, with ablations confirming the necessity of each component. The approach advances interpretable, reliable planning by tying cognitive reasoning to real-world motion constraints, offering a path toward more robust autonomous systems and future work in multi-agent coordination and sensor fusion.

Abstract

Vision-Language-Action (VLA) models in autonomous driving systems have recently demonstrated transformative potential by integrating multimodal perception with decision-making capabilities. However, the interpretability and coherence of the decision process and the plausibility of action sequences remain largely underexplored. To address these issues, we propose AutoDrive-R$^2$, a novel VLA framework that enhances both reasoning and self-reflection capabilities of autonomous driving systems through chain-of-thought (CoT) processing and reinforcement learning (RL). Specifically, we first propose an innovative CoT dataset named nuScenesR$^2$-6K for supervised fine-tuning, which effectively builds cognitive bridges between input information and output trajectories through a four-step logical chain with self-reflection for validation. Moreover, to maximize both reasoning and self-reflection during the RL stage, we further employ the Group Relative Policy Optimization (GRPO) algorithm within a physics-grounded reward framework that incorporates spatial alignment, vehicle dynamic, and temporal smoothness criteria to ensure reliable and realistic trajectory planning. Extensive evaluation results across both nuScenes and Waymo datasets demonstrates the state-of-the-art performance and robust generalization capacity of our proposed method.

Figures (7)

• Figure 1: AutoDrive-R² can effectively achieve planning trajectories across multiple benchmarks compared with other models. Given vehicle's initial visual inputs and language information, AutoDrive-R² achieves comprehensive contextual reasoning for precise trajectory planning. Our model achieves state-of-the-art performance on both nuScenes and Waymo benchmarks.
• Figure 2: Pipeline of AutoDrive-R². We adopt a two-stage training process. The first stage introduce an innovative CoT dataset named nuScenesR²-6K for SFT. The nuScenesR²-6K adopts a four-step logical chain with self-reflection to generate valuable chain-of-thought data. The second stage propose an novel physics-grounded reward framework within the GRPO algorithm for RL, which incorporates spatial alignment, vehicle dynamic, and temporal smoothness for reliable trajectory planning.
• Figure 3: Qualitative comparison of trajectory planning performance across Qwen2.5-VL-7B, EMMA+, and our AutoDrive-R² on the nuScenes dataset. Note that blue lines denote predicted trajectories while green lines represent ground truth trajectories.
• Figure 4: Qualitative comparison of trajectory planning performance across Qwen2.5-VL-7B, EMMA+, and our AutoDrive-R² on the nuScenes dataset. Note that blue lines denote predicted trajectories while green lines represent ground truth trajectories.
• Figure 5: Visualization comparison bewtwee Qwen2.5-VL-7B our AutoDrive-R² on the nuScenes dataset.