IPADS-SAI/MobiAgent

MobiAgent – A Customizable Mobile‑Agent Framework

What it is – MobiAgent is an open‑source system that lets researchers and developers run large‑language‑model‑based “agents” directly on a smartphone (Android or HarmonyOS). It bundles three core pieces:

Component	Role
MobiMind	A family of LLMs (decider, planner, grounder) tuned for mobile tasks.
AgentRR	Agent Record & Replay – a lightweight framework that caches successful action sequences and can replay them to speed up later runs.
MobiFlow	A benchmark that evaluates mobile agents on multi‑step, cross‑app workflows (e.g., searching, purchasing, messaging).

Key Features

On‑device inference – models can be served with vLLM on the phone or via a local server; a quantized 2‑bit/8‑bit version (MobiMind‑1.5‑2B‑W8A8) is provided.
Modular memory systems
- User‑profile memory (Mem0) stored in Milvus vector DB or Neo4j GraphRAG for personalized planning.
- Experience memory that retrieves similar past executions.
- Action memory (AgentRR) that reuses cached action trees.
Multi‑task execution – a runner that can orchestrate complex, cross‑application workflows (e.g., “search on 小红书 → buy on 淘宝 → send result via 微信”).
Unified GUI runner – one‑click configuration for many back‑ends (MobiAgent, UI‑TARS, AutoGLM, Qwen‑VL, Gemini, etc.).
Extensible workflow pipeline – raw‑data cleaning and custom workflow definitions are supported.
Benchmarks & results – the repo includes the MobiFlow benchmark and visual evaluation charts.

Architecture Overview

+-------------------+      +-------------------+      +-------------------+
|   Decider Model   | ---> |   Planner Model   | ---> |   Grounder Model  |
+-------------------+      +-------------------+      +-------------------+
        ^                         ^                         ^
        |                         |                         |
        |   AgentRR (record/replay)   |   Memory modules (profile, experience)
        |                         |                         |
        +-------------------+-----------------------------+
                            |
                        Runner
        (connects to phone via ADB, issues UI actions, records traces)

The diagram in the README (assets/arch.png) visualises the same flow.

Getting Started (quick‑start)

Install the Python environment – create a conda env (python=3.10) and pip install -r requirements.txt (or the simpler requirements_simple.txt if you only want the runner).
Set up the phone – install the ADBKeyboard APK, enable USB debugging, and connect via USB.
Deploy a model – download a MobiMind checkpoint (e.g., MobiMind-1.5-4B) from HuggingFace or ModelScope and serve it with vllm serve … --port <port>.
(Optional) Configure memory – start a Milvus container for vector search or a Neo4j container for GraphRAG, then fill .env with the URLs and model paths.
Run the agent – edit runner/mobiagent/task.json with the tasks you want, then launch:
```
python -m runner.mobiagent.mobiagent \
    --service_ip localhost \
    --decider_port 8000 \
    --planner_port 8002 \
    [--user_profile on] [--use_experience]
```
For multi‑task scenarios use the runner.mobiagent.multi_task.mobiagent_refactored entry point.
Try the Android app – download the pre‑built APK from the releases page (v1.0.1) for a point‑and‑click experience.

Example Use Cases

Personal assistant – retrieve user‑specific recommendations, shop, and send results via messaging apps.
Enterprise automation – automate repetitive mobile workflows (e.g., data entry across internal apps).
Research – benchmark new prompting or planning strategies on the MobiFlow suite; experiment with memory‑augmented agents.

Evaluation

The README includes three result figures (assets/result1.png‑3.png) and a performance chart for AgentRR (assets/result_agentrr.png). These show that the framework achieves higher task‑completion rates and lower latency when using the replay and memory modules compared with a baseline LLM‑only agent.

Release Highlights (chronological)

2025‑08‑30 – Open‑source launch.
2025‑12‑26 – Pure on‑device inference support.
2025‑12‑08 – Release of MobiMind-Reasoning-4B and its AWQ‑quantized variant.
2026‑07‑17 – First proactive on‑device agent system ClawMate, on‑device inference framework MobiInfer, and a quantized 2‑bit model.
Ongoing additions: custom workflows, multi‑task execution, user‑profile memory, mixed‑modal models (Qwen3‑VL), etc.

Citation

If you use MobiAgent in academic work, cite the two arXiv papers linked in the README:

@misc{zhang2025mobiagentsystematicframeworkcustomizable,
  title={MobiAgent: A Systematic Framework for Customizable Mobile Agents},
  author={Cheng Zhang and Erhu Feng and Xi Zhao and Yisheng Zhao and Wangbo Gong and Jiahui Sun and Dong Du and Zhichao Hua and Yubin Xia and Haibo Chen},
  year={2025},
  eprint={2509.00531},
  archivePrefix={arXiv},
  primaryClass={cs.MA},
  url={https://arxiv.org/abs/2509.00531}
}
@misc{liu2025trainingenablingselfevolutionagents,
  title={Beyond Training: Enabling Self‑Evolution of Agents with MOBIMEM},
  author={Zibin Liu and Cheng Zhang and Xi Zhao and Yunfei Feng and Bingyu Bai and Dahu Feng and Erhu Feng and Yubin Xia and Haibo Chen},
  year={2025},
  eprint={2512.15784},
  archivePrefix={arXiv},
  primaryClass={cs.AI},
  url={https://arxiv.org/abs/2512.15784}
}

Where to Find More

Paper: https://arxiv.org/abs/2509.00531 (MobiAgent) and https://arxiv.org/abs/2512.15784 (MobiMem).
Model hub: HuggingFace collection IPADS-SAI/mobimind.
Demo videos – linked in the README via GitHub attachment URLs.
Sub‑module docs – each folder (agent_rr/, runner/, MobiFlow/, etc.) contains its own README.md with deeper usage instructions.

All information above is taken directly from the repository’s README; no external assumptions have been added.