Attacking Motion Planners Using Adversarial Perception Errors

The article by Jonathan Sadeghi et al. discusses the potential for adversarial perception errors in autonomous driving systems. These errors, despite scoring highly on perception quality metrics, can lead to planning failures. The authors demonstrate the creation of such errors using a boundary-attack algorithm and test its effectiveness on two different black-box planners in various driving scenarios using the CARLA simulator. The study concludes that these adversarial perception errors have significant implications for the deployment and testing of autonomous driving systems.

Publication date: 22 Nov 2023
Project Page: Not provided
Paper: https://arxiv.org/pdf/2311.12722