jiagengliu02 / AutoAgent0

# AutoAgent0: An Agentic Runtime for Safe Closed-Loop Driving
Jiageng Liu, Seth Z. Zhao, Zewei Zhou, Alejandro Huerta, Aidan Bayer-Calvert, Yuxin Bao, Haotian Lin, Robert Spataru, Zijun Cui, Yining Hong, Bolei Zhou, Jiaqi Ma
UCLA · Carnegie Mellon University · UT Dallas · Stanford University
Paper | Project Page | Quick Start
https://github.com/user-attachments/assets/11e13673-be87-4c94-b13d-291de7a89fda
## Introduction
**AutoAgent0** is an agentic runtime for safe closed-loop driving. It augments frozen driving experts with independent scene assessment, trajectory verification, and feedback-driven recovery.
- **Scene context refinement** combines complementary perception skills to identify safety-critical actors and refine the context used for verification.
- **Rule-based verification** checks nominal and recovery trajectories before execution.
- **Recovery planning** constructs alternative maneuvers from action primitives and iteratively refines them using verifier feedback when nominal execution becomes unreliable.
## Results
**Fail2Drive.** Driving score (DS) and success rate (SR), averaged over three runs.
| Method | In-distribution DS ↑ | Generalization DS ↑ | Average DS ↑ | Average SR (%) ↑ |
| --- | ---: | ---: | ---: | ---: |
| UniAD | 47.5 | 44.0 | 45.8 | 32.0 |
| + AutoAgent0 | **68.4** | **69.2** | **68.8** | **69.0** |
| SparseDriveV2 | 70.0 | 62.9 | 66.5 | 63.4 |
| + AutoAgent0 | **80.7** | **79.2** | **80.0** | **83.4** |
| BridgeDrive | 91.6 | 81.9 | 86.8 | 85.0 |
| + AutoAgent0 | **91.7** | **85.7** | **88.7** | **90.3** |
## Quick Start
Installation and evaluation
## Citation
 [code]