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

A new benchmark called ARB4WM provides a unified framework for testing the adversarial robustness of world-model agents used in robotics and continuous control, researchers report. The framework evaluates threats across policy, value, and latent-dynamics levels under visual perturbations [1]. World models learn latent dynamics for planning and decision-making and are widely deployed in robotic and agentic engineering control systems [1]. As these systems move into safety-critical settings, understanding their vulnerability to adversarial conditions has become essential, yet existing evaluations have lacked a unified benchmark for testing threats across multiple component levels [1]. ARB4WM addresses that gap by defining five white-box loss objectives targeting the policy, value estimation, and latent-dynamics components of world-model agents [1]. The framework combines those objectives with single-step or multi-step perturbation strategies and temporal attack modes, including full-frame, half-sequence, and sparse-frame exposure [1]. The researchers evaluated four Dreamer-style agents across 20 tasks drawn from MetaWorld and the DeepMind Control Suite [1]. The DeepMind Control Suite is a set of continuous control environments developed by Google DeepMind, the British-American AI research laboratory that has produced systems ranging from AlphaGo to the Gemini family of large language models [5][7]. The study found that attacks targeting value estimation, latent representations, and RSSM dynamics can be as damaging as direct policy disruption [1]. Early or frequent perturbations proved especially harmful, while input-level defenses offered limited recovery under adaptive attacks [1]. The findings arrive amid broader scrutiny of pre-deployment AI safety practices. A recent analysis of 1,178 safety and reliability papers published between January 2020 and March 2025 found that corporate AI research increasingly concentrates on pre-deployment areas such as model alignment and testing and evaluation, while attention to deployment-stage issues like model bias has waned [4]. The ARB4WM authors argue that safety, risk, and reliability assessment for world models should cover multiple component-oriented attack objectives and temporal exposure protocols rather than relying solely on action-space robustness [1]. Source code for the benchmark is publicly available on GitHub [1].