MobileGym: A Lightweight Simulation Platform for Mobile GUI Agent Research
MobileGym serves as a lightweight simulation environment accessible via a browser, designed for mobile GUI agents. It allows for verifiable outcome signals through deterministic evaluations based on structured JSON states. The system facilitates scalable online reinforcement learning with economical parallel rollouts, enabling a single server to manage hundreds of instances, each utilizing approximately 400 MB of memory and requiring a 3-second cold start. The platform incorporates a layered state model alongside a declarative framework for task definitions, enhancing state programmability and task development. Additionally, the MobileGym-Bench offers 416 parameterized task templates, while a unified programmatic judging mechanism provides both deterministic evaluation outcomes and dense RL rewards.
Key facts
- MobileGym is a browser-hosted, lightweight simulation environment.
- It enables verifiable outcome signals via deterministic state-based judging over structured JSON state.
- Supports scalable online RL through low-cost parallel rollouts.
- A single server can host hundreds of parallel instances.
- Each instance uses about 400 MB memory and has a 3-second cold start.
- Features a layered state model and declarative task-definition framework.
- A single programmatic judging mechanism provides deterministic evaluation verdicts and dense RL rewards.
- MobileGym-Bench includes 416 parameterized task templates.
Entities
—