MobileGym

MobileGym is an open-source AI developer tool: a verifiable, highly parallel simulation platform for mobile GUI agent research. It runs an Android-like operating system entirely in the browser, implemented in TypeScript and React, so researchers can train and evaluate agents that operate real-style mobile apps without physical devices. The sandbox ships with 28 simulated apps — including WeChat, Alipay, Xiaohongshu/RED, bilibili, X, Reddit, WeChat Read, China Railway 12306, Tencent Meeting, Spotify, and eBay — and 416 parameterized task templates. Its core advantage is programmatic state verification: structured environment-state snapshots enable reliable pass/fail evaluation (versus a reported 10.2% VLM-judge misjudgment rate) and identical-state replication for reproducible experiments. A live "State Builder" lets researchers inject or patch runtime state without restarting the device — snapshotting, time-travel, and cross-app data injection while the simulator keeps running. The platform is designed for scalable online reinforcement learning, including single-machine batch-parallel GRPO training, and reports a +40.7-point sim-to-real improvement. Researchers can connect the built-in demo model or their own OpenAI-compatible vision endpoint to run agent tasks. MobileGym is aimed at ML engineers and robotics/agent researchers who need reproducible, verifiable environments for training and benchmarking mobile GUI agents, and its code and paper are openly available.

ML engineers and agent researchers use MobileGym to train and benchmark mobile GUI agents in a reproducible, verifiable browser sandbox.