惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

J
Java Code Geeks
MongoDB | Blog
MongoDB | Blog
B
Blog
博客园 - Franky
博客园 - 三生石上(FineUI控件)
A
About on SuperTechFans
N
Netflix TechBlog - Medium
MyScale Blog
MyScale Blog
阮一峰的网络日志
阮一峰的网络日志
美团技术团队
Vercel News
Vercel News
云风的 BLOG
云风的 BLOG
WordPress大学
WordPress大学
S
SegmentFault 最新的问题
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
宝玉的分享
宝玉的分享
小众软件
小众软件
P
Proofpoint News Feed
aimingoo的专栏
aimingoo的专栏
freeCodeCamp Programming Tutorials: Python, JavaScript, Git & More
月光博客
月光博客
酷 壳 – CoolShell
酷 壳 – CoolShell
D
DataBreaches.Net
量子位

DEV Community

Authentication Security Deep Dive: From Brute Force to Salted Hashing (With Java Examples) Why AI Systems Don’t Fail — They Drift Spilling beans for how i learn for exam😁"Reinforcement Learning Cheat Sheet" I Replaced Chrome with Safari for AI Browser Automation. Here's What Broke (and What Finally Worked) How Python Borrows Other People's Work The $40 Architecture: Processing 1 Billion API Requests with 99.99% Uptime Vibe Coding: A Workflow Guide (From Zero to SaaS) Most webhook security guides protect the wrong side. The scary part is delivery. Headless CMS for TanStack Start: Build a Blog with Cosmic EU Age Verification App "Hacked in 2 Minutes" — What Actually Happened Comfy Cloud’s delete function does not actually remove files Running AI Models on GPU Cloud Servers: A Beginner Guide Event-driven media intelligence with AWS Step Functions and Bedrock I scored 500 AI prompts across 8 quality dimensions — here's what broke How to Call Google Gemini API from Next.js (Free Tier, No Backend Needed) The Portal Protocol: Reclaiming Human Connection in the Age of AI How to Fix Your Team's Scattered Knowledge Problem With a Self-Hosted Forum Intro to tc Cloud Functors: A Graph-First Mental Model for the Modern Cloud Designing Multi-Tenant Backends With Both Ownership and Team Access I Built a Neumorphic CSS Library with 77+ Components — Here's What I Learned PostgreSQL Performance Optimization: Why Connection Pooling Is Critical at Scale Cómo construí un SaaS multi-rubro para gestionar expensas en Argentina con FastAPI + Vue 3 🚀 I Built an Ethical Hacking Scanner Tool – Open Source Project I Replaced /usage and /context in Claude Code With a Single Statusline A Pythonic Way to Handle Emails (IMAP/SMTP) with Auto-Discovery and AI-Ready Design I Collected 8.9 Million Polymarket Price Points — Here's What I Found About How Markets Really Move EcoTrack AI — Carbon Footprint Tracker & Dashboard Everyone's Using AI. No One Agrees How. 5 self-hosted ebook managers worth trying in 2026 Building Your First AI Agent with LangChain: From Chatbot to Autonomous Assistant
After burning through tens of billions of tokens, I built...
rui hao · 2026-05-29 · via DEV Community

rui hao

After burning through tens of billions of tokens, I built an Android-like OS that runs entirely in the browser

The title is a bit clickbaity, but it is not that far from what actually happened.

Over the past few months, we built MobileGym: a fully browser-based, Android-like simulation environment implemented in TypeScript + React.

It currently includes 28 simulated apps, including WeChat, Alipay, Xiaohongshu/RED, bilibili, X, Reddit, WeChat Read, China Railway 12306, Tencent Meeting, Spotify, and eBay, plus system apps such as Home, Settings, Contacts, Messages, Photos, Calendar, Files, and Browser.

The system supports Xiaomi theme packs and custom widgets. We also reimplemented a number of Android-like system mechanisms directly in the browser, including the Activity stack, Intents, gesture navigation, back handling, and soft keyboard behavior.

MobileGym was originally built for GUI agent research, but it is also open for anyone who wants to play with it, study Android-like UI/system mechanisms, or fork the code and build something else on top of it.

Online demo: https://mobilegym.dev
GitHub: https://github.com/Purewhiter/mobilegym

Features

⚡ Lightweight and highly concurrent
A single MobileGym instance uses only around 400 MB of memory, compared with roughly 4–10 GB for a typical Android emulator. A single server can run hundreds or even thousands of environment instances in parallel.

🧪 416 task templates
The task templates are parameterized, so they can generate an effectively unlimited number of task instances. Evaluation is deterministic and finishes in milliseconds, without relying on LLM-as-a-judge.

🌉 Sim-to-real transfer that actually works
In our tests, models trained with GRPO-style reinforcement learning in the simulated environment transferred more than 95% of their gains to real devices.

🧩 Easy to extend
MobileGym is designed to be extensible. Adding a new app only requires creating a folder and a manifest file. Adding a new task only requires writing a Python class, and the shortest tasks can be implemented in as little as three lines of code.

🛡️ Fully sandboxed, with no real-world consequences
MobileGym does not connect to real services, transfer real money, or send real messages. You can safely click around without worrying about side effects.

Although the project started as infrastructure for GUI agent training and evaluation, it ended up becoming a fairly complete browser-based Android-like playground.
Would love to hear feedback from anyone who finds this useful, interesting, or just fun to play with.