Purewhiter / mobilegym

What is mobilegym?

MobileGym is a browser-based simulation environment for training and evaluating mobile GUI agents. Built with TypeScript and React, it replicates 28 popular apps like Alipay, WeChat, and Bilibili inside a virtual Android-like shell. The platform exposes the entire application state as structured JSON, allowing programmatic verification of agent actions instead of relying on slow, error-prone vision models. Researchers get a sandbox where agents can transfer money, send messages, and navigate complex flows without touching real accounts or funds.

Why is it gaining traction?

The key differentiator is deterministic verification. Traditional benchmarks fall back on VLM judges that misjudge outcomes roughly 10% of the time. MobileGym's state-based judges run in sub-milliseconds and catch unintended side effects like accidentally following a user. The platform also scales horizontally: you can run 256 parallel simulator instances on a single server using under 10% CPU, enabling fast iteration cycles for reinforcement learning experiments. Perhaps most compellingly, the team demonstrated that 95.1% of training gains transfer to a real Redmi Note 12 Turbo, bridging the sim-to-real gap that plagues most agent benchmarks.

Who should use this?

Researchers benchmarking mobile GUI agents will find the most value here. The pre-built task suite with 416 templates covers diverse scenarios, and the programmatic judge system eliminates the overhead of building custom evaluation pipelines. RL practitioners interested in online training for vision-language models on mobile tasks can leverage the parallel rollout infrastructure. App developers prototyping automation flows might also use it as a low-stakes testing environment, though this is secondary to the research focus.

Verdict

With a credibility score of 0.949999988079071% and only 29 stars, MobileGym is early-stage but backed by a detailed arXiv paper and validated sim-to-real transfer numbers. The documentation is thorough, the modular app architecture makes extension straightforward, and the dual Apache 2.0 / CC BY-NC 4.0 licensing separates reusable code from research content. Worth evaluating now if you're serious about mobile agent research, but watch for maturity as the community grows.