Show HN: I built an Android-like OS that runs in the browser (mobilegym.dev)

A new simulation platform called MobileGym has been developed, allowing for the online reinforcement learning (RL) training and deterministic evaluation of mobile GUI agents directly in a web browser. This platform supports 28 mobile applications, significantly enhancing the ability to conduct research in environments involving real-world daily apps, which have historically posed challenges in traditional real-device testing due to issues like state untraceability and limited accessibility. The platform operates using structured JSON states, enabling quick snapshots and restoration of environments, which permits efficient multi-instance simulations on a single machine. The significance of MobileGym lies in its ability to overcome critical barriers in mobile app interaction research by providing a reliable framework for benchmarking and training agents on real-world apps. Notably, it has yielded impressive results, with a reinforcement-fine-tuning approach boosting simulation success rates (SR) by 12.8% and 40.7% retention in real-device performance. Additionally, MobileGym operates with a more efficient resource footprint compared to traditional emulation, making it a compelling solution for researchers looking to advance the capabilities and training methodologies of mobile AI agents.