ARB4WM: An Adversarial Robustness Benchmark for World Models in Continuous Control

The paper introduces ARB4WM, a benchmark designed to evaluate the adversarial robustness of world models in continuous control environments. It defines five white-box loss objectives across policy, value, and latent dynamics, and assesses their impact on four Dreamer-style agents across 20 tasks from MetaWorld and the DeepMind Control Suite. The findings indicate that various attack strategies can significantly undermine performance, highlighting the need for comprehensive robustness assessments that go beyond traditional action-space evaluations, which is crucial for practitioners developing safety-critical AI systems.