Skip to content
InfoResearchPreprint

STA: A Split Temporal Adversary Method for Evaluating Robustness of Multimodal RL-based Driving Policies

Published
Record updated
View JSON

Summary

Researchers propose the Split Temporal Adversary (STA), a method that evaluates the robustness of multimodal reinforcement learning (MRL) driving policies by attacking their temporal alignment across sensor modalities. STA delays observations via historical states and jointly optimizes attack timing and modality selection under budget constraints. In CARLA simulator experiments, STA more than doubles the collision rate of the strongest baseline against PPO while using 17.8% fewer attack attempts, and it also generalizes to a VLM-based driving policy.