CoReVLA: A Dual-Stage End-to-End Autonomous Driving Framework for Long-Tail Scenarios via Collect-and-Refine

TL;DR

The paper tackles the difficulty of long-tail, safety-critical driving scenarios by introducing CoReVLA, a dual-stage Collect-and-Refine framework that first fine-tunes on driving QA data, then collects driver takeover data in a CAVE HITL environment and refines behavior via Direct Preference Optimization. This continual learning loop enables the model to learn from human preferences in rare events, improving robustness and safety. Open-loop QA tests show improved understanding and reasoning, while closed-loop Bench2Drive results demonstrate state-of-the-art Driving Score and Success Rate in long-tail conditions, with evidence of continual improvement through past takeovers. The approach offers a practical pipeline for data collection, refinement, and continual learning in autonomous driving, with open-source code and datasets available.

Abstract

Autonomous Driving (AD) systems have made notable progress, but their performance in long-tail, safety-critical scenarios remains limited. These rare cases contribute a disproportionate number of accidents. Vision-Language Action (VLA) models have strong reasoning abilities and offer a potential solution, but their effectiveness is limited by the lack of high-quality data and inefficient learning in such conditions. To address these challenges, we propose CoReVLA, a continual learning end-to-end autonomous driving framework that improves the performance in long-tail scenarios through a dual-stage process of data Collection and behavior Refinement. First, the model is jointly fine-tuned on a mixture of open-source driving QA datasets, allowing it to acquire a foundational understanding of driving scenarios. Next, CoReVLA is deployed within the Cave Automatic Virtual Environment (CAVE) simulation platform, where driver takeover data is collected from real-time interactions. Each takeover indicates a long-tail scenario that CoReVLA fails to handle reliably. Finally, the model is refined via Direct Preference Optimization (DPO), allowing it to learn directly from human preferences and thereby avoid reward hacking caused by manually designed rewards. Extensive open-loop and closed-loop experiments demonstrate that the proposed CoReVLA model can accurately perceive driving scenarios and make appropriate decisions. On the Bench2Drive benchmark, CoReVLA achieves a Driving Score (DS) of 72.18 and a Success Rate (SR) of 50%, outperforming state-of-the-art methods by 7.96 DS and 15% SR under long-tail, safety-critical scenarios. Furthermore, case studies demonstrate the model's ability to continually improve its performance in similar failure-prone scenarios by leveraging past takeover experiences. All codea and preprocessed datasets are available at: https://github.com/FanGShiYuu/CoReVLA

Figures (5)

• Figure 1: Overview of the proposed CoReVLA framework.
• Figure 2: Human-in-the-loop testing and takeover data collection in the CAVE platform.
• Figure 3: Comparison of model responses in open-loop scene understanding and decision-making.
• Figure 4: Comparison of CoReVLA driving trajectories before and after behavior refinement.
• Figure 5: Scenario comparison illustrating CoReVLA’s behavior refinement and generalization across CAVE and Bench2Drive.