GUI agents deployed on Android devices frequently encounter dynamic anomalies such as unexpected pop‑ups and action misuse, yet existing benchmarks lack systematic robustness evaluation. We therefore introduce AnTrap, which injects dynamic perturbations into agents' execution trajectories. A taxonomy is proposed that organizes real‑world anomalies into four layers—State, Thinking, Action, and Round—with ten fine‑grained subcategories, and a construction pipeline that preserves task solvability while delivering realistic adversarial conditions. Evaluation of 16 leading GUI models shows universal vulnerability: even the strongest models suffer notable performance drops under dynamic anomalies. We further conduct GRPO training in both original and adversarial environments to separate anomalies learnable from the environment from those limited by reasoning bottlenecks. Results indicate that single‑step traps at the State and Action layers can be largely mitigated via adversarial reinforcement learning, whereas deep contextual traps such as state deadlock expose intrinsic reasoning limitations that cannot be overcome merely by training in trap‑laden environments.
Blogger's Review: AnTrap offers a rigorous framework for probing GUI agents' robustness, highlighting that while adversarial RL can address superficial, single‑step anomalies, deeper contextual failures remain a fundamental challenge, urging future work to rethink agent reasoning architectures.