Top AI Product

MobileGym is a browser-hosted simulation platform for training and evaluating mobile GUI agents — the agents that tap, swipe, and type through phone apps. Its bet is that you don’t need to replicate proprietary app backends to train an agent; you need fidelity in the interaction and a way to verify outcomes cheaply.

## Verifiable state, massively parallel

The whole environment state is captured as structured JSON that can be configured, forked, and compared — which gives deterministic, state-based judging instead of brittle screenshot matching. That same design makes it cheap to parallelise: one server hosts hundreds of instances at roughly 400MB each with about a 3-second cold start. The companion MobileGym-Bench ships 416 parameterised task templates across 28 apps with deterministic judges.

## Why it matters

Mobile GUI agents have been stuck on a data problem: real devices are slow and unverifiable, and most simulators can’t produce trustworthy reward signals for reinforcement learning. MobileGym’s sim-to-real result is the proof point — GRPO on Qwen3-VL-4B gains 12.8 points on the test set, and real-device execution retains 95.1% of that simulation-trained gain. Cheap parallel rollouts plus rewards that transfer to hardware is exactly what scalable online RL for phone agents needs.