Summary

GELab-Zero addresses mobile GUI agent development complexities with an open-source, plug-and-play infrastructure and a local 4B agent model. It enables developers and enterprises to accelerate innovation, ensure data privacy, and eliminate cloud dependencies by abstracting intricate deployment requirements.

How It Works

The project provides a self-contained, locally deployable inference infrastructure and a 4B GUI agent model, managing challenges like multi-device ADB, dependencies, and inference services. Key features include a unified one-click deployment pipeline, multi-phone task distribution, and three agent modes: ReAct, multi-agent collaboration, and scheduled tasks, prioritizing local execution on consumer hardware for privacy.

Quick Start & Requirements

• Installation: Requires Python 3.12+ (Miniforge recommended), an LLM inference environment (Ollama/vLLM), and an Android device with developer mode/USB debugging. Setup involves cloning the repo, installing dependencies (pip install -r requirements.txt), configuring the LLM backend, and connecting the Android device via ADB.
• Prerequisites: Python 3.12+, ADB tool, Ollama/vLLM, Android device.
• Links:
  • Repository: https://github.com/stepfun-ai/gelab-zero
  • Ollama: https://ollama.com/

Highlighted Details

• AndroidDaily Benchmark: Introduces a novel benchmark for real-world mobile lifestyle apps, featuring static (3146 actions) and end-to-end (235 tasks) evaluations.
• Performance: GELab-Zero-4B-preview achieves 0.734 accuracy on the AndroidDaily static benchmark and 75.86% success on end-to-end mobile tasks.
• Local Deployment: Supports 4B models on consumer hardware, balancing latency and privacy.
• Agent Modes: Offers ReAct loops, multi-agent collaboration, and scheduled task execution.

Maintenance & Community

Developed by the StepFun-GELab team. A WeChat group is mentioned for communication, but specific contributor details, sponsorships, or a public roadmap are absent from the README. Contact: tankaijun@stepfun.com.

Licensing & Compatibility

The repository's license is not explicitly stated in the README, creating ambiguity for commercial use, derivative works, and integration into closed-source projects.

Limitations & Caveats

The primary model, GELab-Zero-4B-preview, is marked as "preview," indicating potential instability. While aiming for one-click deployment, the setup requires integrating multiple technical environments (Python, LLM inference, ADB), which may challenge users unfamiliar with advanced command-line operations and system configuration.