Purewhiter/mobilegym

MobileGym runs a real Android environment inside a browser, then wraps it in infrastructure for scalable online reinforcement learning and verifiable evaluation of GUI agents. The two things that matter here: first, no device farm — the simulator is browser-hosted and horizontally scalable, which cuts the cost and coordination overhead that usually makes mobile agent research painful. Second, the verifiable evaluation layer means rollout correctness is checkable rather than eyeballed, which matters the moment you are comparing agent checkpoints or debugging reward hacking. The project is research-grade — the README is dense and leans on academic framing, and getting a training run going will take real setup time. But if your team is building or benchmarking mobile GUI agents and you have been cobbling together emulator fleets, this is worth a serious look. The parallel RL training path is the differentiator over simpler emulator wrappers. -> Best for: ML researcher or AI engineer working on mobile GUI or device-control agents