MobileGym: A Verifiable and Highly Parallel Simulation Platform for Mobile GUI Agent Research

Dingbang Wu , Rui Hao , Haiyang Wang , Shuzhe Wu , Han Xiao , Zhenghong Li , Bojiang Zhou , Zheng Ju , Zichen Liu , Lue Fan , Zhaoxiang Zhang

🏛 Institutions: CASIA , PKU , CUHK
📅 Date: May 25, 2026
📑 Publisher: arXiv
💻 Env: Mobile
🔑 Keywords: benchmark framework reinforcement learning simulation platform sim-to-real MobileGym

TLDR

MobileGym is a browser-hosted Android-like simulation platform for mobile GUI agent research that represents full environment state as structured JSON, enabling deterministic state-based judging, snapshot/reset/fork, side-effect detection, and highly parallel rollouts for online RL. Its MobileGym-Bench provides 416 parameterized task templates over 28 apps, and GRPO training on Qwen3-VL-4B-Instruct improves a 256-task test set by +12.8 points with 95.1% sim-to-real gain retention on a real-device subset.

Open paper arXiv Report issue