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

Zihao Sheng1,*, Zilin Huang1,*, Yansong Qu2, Yue Leng3, Sruthi Bhavanam3, Sikai Chen1,
1University of Wisconsin-Madison,  2Purdue University,  3Google
*Indicates Equal Contribution, Corresponding Author
arXiv Video Code (coming soon)

TL;DR: To the best of our knowledge, CurricuVLM is the first work to utilize VLMs for dynamic curriculum generation in closed-loop autonomous driving training.

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.

Motivation

Descriptive Alt Text

Comparison of training strategies for autonomous driving agents. (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.

Framework

Descriptive Alt Text

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.

Visualization

Case 1: Unprotected Right Turn

Descriptive Alt Text

Before training with CurricuVLM.

Descriptive Alt Text

After training with CurricuVLM.

A blue car makes an unprotected right-turning, the untrained agent fails to maintain a safe distance and collides with the turning vehicle. After training with CurricuVLM, the agent learns to proactively adjust its speed and lateral position, creating a larger safety margin while maintaining efficient progress.


Case 2: Merge

Descriptive Alt Text

Before training with CurricuVLM.

Descriptive Alt Text

After training with CurricuVLM.

The agent encounters a merging situation where untrained behavior results in collisions with other vehicles. After training with CurricuVLM, the agent develops more sophisticated merging behavior, learning to select a safer lane and dynamically adjust its speed to avoid potential conflicts with surrounding vehicles.


Case 3: Front Car Deceleration

Descriptive Alt Text

Before training with CurricuVLM.

Descriptive Alt Text

After training with CurricuVLM.

This case showcases the agent’s improved response to sudden speed changes from leading vehicles. While the untrained agent fails to react appropriately to a decelerating vehicle, the trained agent successfully executes a safe lane change maneuver to maintain smooth traffic flow.


Case 4: Abrupt Lane Intrusion

Descriptive Alt Text

Before training with CurricuVLM.

Descriptive Alt Text

After training with CurricuVLM.

The agent learns to anticipate and respond to potential lane change conflicts. When confronted with an adjacent vehicle’s lane change, the trained agent demonstrates sophisticated decision-making by accelerating and shifting to a safer lane while maintaining appropriate following distances.


Case 5: Cut-in Lane Collision

Descriptive Alt Text

Before training with CurricuVLM.

Descriptive Alt Text

After training with CurricuVLM.

In this scenario, a vehicle from an adjacent lane cuts in abruptly in front of the red car, leading to a collision due to insufficient time and space for the red car to react. After training with CurricuVLM, the agent learns to anticipate such abrupt cut-ins, proactively adjusting its speed and maintaining a safer following distance.

BibTeX

@article{sheng2025curricuvlm,
  title={CurricuVLM: Towards Safe Autonomous Driving via Personalized Safety-Critical Curriculum Learning with Vision-Language Models},
  author={Sheng, Zihao},
  journal={arXiv preprint arXiv:2502.15119},
  year={2025}
}