MobileGym: A Verifiable and Highly Parallel Simulation Platform for Mobile GUI Agent Research
Dingbang Wu, Rui Hao, Haiyang Wang, Shuzhe Wu, Han Xiao, Zhenghong Li, Bojiang Zhou, Zheng Ju, Zichen Liu, Lue Fan, Zhaoxiang Zhang
Key claim
MobileGym enables scalable online reinforcement learning for mobile apps.
MobileGym is a new environment designed for mobile applications that allows for high interaction fidelity and scalable reinforcement learning. It provides structured evaluation and rewards, leading to a notable performance improvement in real-device execution. The key result shows a +12.8 percentage point gain on a test set, indicating its effectiveness.
MobileGym introduces a novel environment for mobile app interaction and scalable online reinforcement learning.
The claims are supported by a comprehensive set of task templates and a case study demonstrating significant performance gains.
Deep reliability assessment
The methodology supports the creation of a lightweight, browser-hosted simulation environment for mobile GUI agents, but the claim of interaction fidelity without replicating proprietary backends may be overclaimed as it does not fully capture backend dynamics.
Reproducibility
yes, the project page is provided: https://mobilegym.github.io.
Discussion questions
- How does MOBILEGYM ensure interaction fidelity without replicating proprietary backends?
- What are the practical implications of using MOBILEGYM for training mobile GUI agents in real-world applications?
- What evidence would falsify the claim that MOBILEGYM's training gains transfer effectively to real-device execution?
Key figure
Figure 1 shows example screens from MOBILEGYM, highlighting its configurable and sandboxed capabilities.