CAT: Closed-loop Adversarial Training for
Safe End-to-End Driving

7th Annual Conference on Robot Learning (CoRL 2023)

Linrui Zhang 1 , Zhenghao Peng 2 , Quanyi Li 3 , Bolei Zhou 2
1 Tsinghua University , 2 University of California, Los Angeles , 3 University of Edinburgh
| Code | Paper

1min Spotlight Video

This video includes audio.

Overview

Fig. 1: Overview of the proposed method.

This work introduces the Closed-loop Adversarial Training (CAT) framework for safe end-to-end driving. CAT imports normal driving scenarios from real-world driving logs and then generates safety-critical counterparts as adversarial training environments tailored to the current driving policy. The agent continuously learns to address emerging challenges and improves risk awareness in a closed-loop pipeline. After training, the agent can achieve superior driving safety in both log-replay (+32.5%) and safety-critical traffic scenarios (+35%).

Fig. 2: Illustration of Factorized Safety-Critical Resampling.

One crucial component of our framework is a novel factorized safety-critical resampling technique that efficiently turns log-replay driving scenarios into safety-critical ones during training.

Specifically, we cast the safety-critical traffic generation as the risk-conditioned Bayesian probability maximization problem and then decompose it into the multiplication of standard motion forecasting sub-problems. Thus, we can utilize off-the-shelf motion forecasting models as the learned prior to generate adversarial scenarios with high fidelity, diversity, and efficiency. Compared to previous adversarial traffic generation methods, the proposed technique obtains a competitive attack success rate while significantly reducing the computational cost (generally less than 1s), making the CAT framework effective and efficient for closed-loop end-to-end driving policy training.

CAT-generated Adversarial Scenarios

Case 1: In the raw scenario, the blue car keeps going straight. After CAT generation, it turns left at the intersection.

Raw scenario
Generated scenario



Case 2: In the raw scene, the blue car drives inside the roundabout. After CAT generation, it runs out of the roundabout.

Raw scenario
Generated scenario



Case 3: In the raw scene, the blue car changes the lane gently. After CAT generation, it suddenly changes the lane.

Raw scenario
Generated scenario


Case 4: In the raw scene, the blue car drive normally on the highway. After CAT generation, it brakes suddenly due to a malfunction.

Raw scenario
Generated scenario

CAT Enhances AI Driving Safety

Case 1: A car makes an unprotected left turn at an intersection. The CAT agent learns to stay away from potentially dangerous vehicles.

Before CAT
After CAT


Case 2: The leading car slows down. The CAT agent learns to change its lane and overtake that vehicle.

Before CAT
After CAT


Case 3: The leading car cuts into the lane suddenly. The CAT agent learns to yield and change its lane ahead of time.

Before CAT
After CAT


Case 4: Two vehicles traveling in opposite directions meet. The CAT agents learns to pass by.

Before CAT
After CAT

Reference

@inproceedings{zhang2023cat,
  title={CAT: Closed-loop Adversarial Training for Safe End-to-End Driving},
  author={Zhang, Linrui and Peng, Zhenghao and Li, Quanyi and Zhou, Bolei},
  booktitle={7th Annual Conference on Robot Learning},
  year={2023}
}


Acknowledgement

This work was supported by the National Science Foundation under Grant No. 2235012 and the Cisco Faculty Award.