<img src="./assets/logo.png" alt="RoboClaw logo" width="28"> RoboClaw: An Agentic Framework for Scalable Long-Horizon Robotic Tasks

<a href="https://roboclaw-agibot.github.io/"> <img src="https://img.shields.io/badge/Project-Website-green" alt="Project Website"> </a> <a href="https://arxiv.org/abs/2603.11558"> <img src="https://img.shields.io/badge/arXiv-2603.11558-b31b1b" alt="arXiv"> </a> <img src="https://img.shields.io/badge/Status-Ready-green" alt="Status: Paper assets only"> <img src="https://img.shields.io/badge/Open--Source-Released-blue" alt="Open-source release plan"> Collaborating Institutions <img src="./assets/Agi-logo.png" alt="AgiBot" height="104">          <img src="./assets/ScaleLab-logo.png" alt="ScaleLab" height="104">

Website | arXiv

RoboClaw is an agentic robotics framework for long-horizon manipulation. It uses a vision-language model as the high-level controller and keeps the same agent in the loop during data collection, policy learning, and deployment.

Instead of treating those stages as separate systems, RoboClaw lets the agent reason over context, choose skills, monitor execution, and feed deployment experience back into training.

RoboClaw teaser

🎉 The public codebase has been released!

✨ Highlights

One agent loop across collection, training, and deployment
Vision-language reasoning for subtask selection and execution monitoring
Entangled Action Pairs (EAP) for self-resetting data collection
Failure recovery through retrying, replanning, or human escalation when needed
Real-world evaluation on long-horizon manipulation tasks with the Agibot G01 platform

📊 Main Results

On long-horizon tasks, RoboClaw improves success rate by 25% over baseline pipelines
Across the full robot lifecycle, it reduces human time investment by 53.7%
For the same amount of collected data, a manual pipeline requires about 2.16x more human time
During rollout, the manual baseline needs about 8.04x more human intervention
Forward policy success rates improve steadily across all four tested subtasks as rollout iterations increase

🧠 What RoboClaw Does

RoboClaw keeps the robot inside a single closed loop:

During data collection, the agent observes the scene, selects the current subtask, and triggers policy execution through MCP-style tools
For each manipulation skill, RoboClaw pairs a forward behavior with an inverse reset behavior through Entangled Action Pairs (EAP), enabling self-resetting collection loops
During deployment, the same agent monitors progress, switches skills when needed, retries or replans after failures, and escalates to humans only when recovery is unreliable or safety constraints are reached
Execution trajectories are fed back into training, so deployment also becomes a source of new experience

🧪 Experimental Scenarios

We evaluate RoboClaw in several real-world settings:

bedroom vanity table organization
kitchen shelf organization
study desk organization
convenience-store shelf retrieval

The single-skill evaluations cover four representative manipulation tasks:

body lotion placement
primer placement with drawer closing
lipstick insertion
tissue wipe

🚀 Quick Start Guide

First, clone the repository:

git clone https://github.com/RoboClaw-Robotics/RoboClaw.git ~/RoboClaw
cd ~/RoboClaw

Then, prepare the .env file in the project root:

cd ~/RoboClaw
cp .env.example .env

Then fill in the tokens and credentials in .env:

FEISHU_APP_ID=cli_xxx
FEISHU_APP_SECRET=xxx
FEISHU_VERIFICATION_TOKEN=xxx
FEISHU_EVENT_RECEIVER=long_connection
OPENAI_API_KEY=xxx

Configuration details:

FEISHU_APP_ID: Feishu application App ID (optional)
FEISHU_APP_SECRET: Feishu application App Secret (optional)
FEISHU_VERIFICATION_TOKEN: Event verification token (optional)
FEISHU_EVENT_RECEIVER=long_connection: Enables long-connection mode (optional)
OPENAI_API_KEY: OpenAI API key (required)

After updating .env, initialize the project:

cd ~/RoboClaw
make init

🔧 Common Commands

Start the TUI: make run_tui
Start the GUI: make run_gui

📄 Paper

Title: RoboClaw: An Agentic Framework for Scalable Long-Horizon Robotic Tasks
Website: https://roboclaw-agibot.github.io/
PDF: assets/paper.pdf
arXiv: arXiv

🚧 Open-Source Release

The current open-source release is primarily validated on the Agibot G01 platform.

For complete usage (including VLA model deployment), the corresponding .whl package is available upon request via private message.

📚 Citation

If RoboClaw is useful for your research, please cite:

@misc{li2026roboclaw,
  title={RoboClaw: An Agentic Framework for Scalable Long-Horizon Robotic Tasks},
  author={Ruiying Li and Yunlang Zhou and YuYao Zhu and Kylin Chen and Jingyuan Wang and Sukai Wang and Kongtao Hu and Minhui Yu and Bowen Jiang and Zhan Su and Jiayao Ma and Xin He and Yongjian Shen and Yangyang and Guanghui Ren and Maoqing Yao and Wenhao Wang and Yao Mu},
  year={2026},
  eprint={2603.11558},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2603.11558}, 
}

RoboClaw

Install / Use

README