RealMobile — Model Trajectories Agent execution trajectories on the RealMobile benchmark — a real device benchmark for mobile GUI agents. Each trajectory is one model attempting one task on a physical Android phone driving live Chinese applications (Bilibili, Douyin, Xiaohongshu, Weibo, Taobao, QQ Music, Ctrip, AMap, Hema, …), recorded step by step: screenshot, UI hierarchy XML, the action taken, and the model's reasoning. This dataset accompanies the Xiaomi GUI 0 technical report. It is intended for qualitative inspection, error analysis, and reproducible re scoring — not for training. 🌐 Benchmark site: https://seerray lab.github.io/Xiaomi GUI 0/realmobile/ 💻 Scoring code: https://github.com/SeerRay Lab/Xiaomi GUI 0/tree/main/realmobile 📄 Technical report: https://arxiv.org/abs/2606.31410 Directory Structure Trajectory Format task.json Each trajectory folder contains a task.json describing the task and every step the agent took: Per step files File Meaning N.png Full resolution screenshot at step N (LFS) N.jpg Compressed screenshot at step N (LFS) N.xml Android UI hierarchy tree at step N ( uiautomator dump ) N.json Action metadata at step N Usage Load a single trajectory: Down…
Runs entirely in your browser via DuckDB-Wasm — this dataset's real data file is loaded once, then queried locally. Nothing is sent to a server.
We use cookies for essential functionality and analytics. You can accept or reject analytics cookies.Cookie policy