CurricuVLM: Towards Safe Autonomous Driving via Personalized Safety-Critical Curriculum Learning with Vision-Language Models

TL;DR

CurricuVLM addresses the challenge of safety in autonomous driving by integrating Vision-Language Models with GPT-4o-based reasoning to create personalized, safety-critical curricula. It forms a closed-loop pipeline where VLM-derived observations and GPT-4o insights drive scenario generation via DenseTNT priors and dynamic scheduling, enabling RL agents to learn safer policies. Extensive experiments on the Waymo Open Motion Dataset show CurricuVLM outperforms traditional RL, Safe RL, imitation learning, and closed-loop curriculum methods across regular and safety-critical scenarios, while remaining compatible with multiple RL algorithms. The framework offers a general, scalable approach to enhancing AV safety training by tailoring curricula to evolving agent weaknesses.

Abstract

Ensuring safety in autonomous driving systems remains a critical challenge, particularly in handling rare but potentially catastrophic safety-critical scenarios. While existing research has explored generating safety-critical scenarios for autonomous vehicle (AV) testing, there is limited work on effectively incorporating these scenarios into policy learning to enhance safety. Furthermore, developing training curricula that adapt to an AV's evolving behavioral patterns and performance bottlenecks remains largely unexplored. To address these challenges, we propose CurricuVLM, a novel framework that leverages Vision-Language Models (VLMs) to enable personalized curriculum learning for autonomous driving agents. Our approach uniquely exploits VLMs' multimodal understanding capabilities to analyze agent behavior, identify performance weaknesses, and dynamically generate tailored training scenarios for curriculum adaptation. Through comprehensive analysis of unsafe driving situations with narrative descriptions, CurricuVLM performs in-depth reasoning to evaluate the AV's capabilities and identify critical behavioral patterns. The framework then synthesizes customized training scenarios targeting these identified limitations, enabling effective and personalized curriculum learning. Extensive experiments on the Waymo Open Motion Dataset show that CurricuVLM outperforms state-of-the-art baselines across both regular and safety-critical scenarios, achieving superior performance in terms of navigation success, driving efficiency, and safety metrics. Further analysis reveals that CurricuVLM serves as a general approach that can be integrated with various RL algorithms to enhance autonomous driving systems. The code and demo video are available at: https://zihaosheng.github.io/CurricuVLM/.

Figures (13)

• Figure 1: Comparison of different approaches for safety-critical scenario integration in autonomous driving. (a) Pre-generated safety-critical scenarios as static training augmentation, which fails to adapt to the agent's evolving capabilities; (b) Conventional curriculum learning with rule-based scenario selection, which lacks personalization to individual learning bottlenecks; (c) Our proposed approach using VLMs as observant mentors and curriculum designers for dynamic and personalized training.
• Figure 2: Overview of CurricuVLM framework. (a) RL environment provides state observations and receives control actions from the agent; (b) Safety-critical event analysis module employs VLM for visual understanding and GPT-4o for behavioral pattern analysis; (c) Personalized curriculum adaptation generates safety-critical scenarios by optimizing background vehicle trajectories; (d) Dynamic curriculum scheduling mechanism adaptively integrates generated scenarios into the training process.
• Figure 3: Experiment environment and evaluation scenarios. (a) MetaDrive simulator with 3D render and top-down view; (b) Representative HD map layouts from Waymo Open Motion Dataset; (c) Regular scenarios with original background vehicle trajectories; (d) Generated safety-critical scenarios with challenging vehicle interactions.
• Figure 4: Performance comparison with traditional RL baselines in both regular and safety-critical test scenarios.
• Figure 5: Performance comparison with safe RL baselines in both regular and safety-critical test scenarios.

Theorems & Definitions (4)

• Definition 1: Driving Scenario
• Definition 2: Curriculum
• Theorem 1
• proof