ARMADA: Autonomous Online Failure Detection and Human Shared Control Empower Scalable Real-world Deployment and Adaptation

Shanghai Jiao Tong University; Shanghai Innovation Institute; Noematrix Ltd.

Wenye Yu, Jun Lv, Zixi Ying, Yang Jin, Chuan Wen, Cewu Lu

Table of Contents

• 🔥 Highlights
• 🛠️ Installation
• 📷 Data Collection
• 🤖 Policy Training
• 🌐 Multi-Robot Deployment
• 📝 TODO
• ✍️ Citation
• 🪪 License
• 🙏 Acknowledgement

🔥 Highlights

• Our failure detector, FLOAT, achieves nearly 95% accuracy in four real-world tasks, improving the SOTA online failure detection approaches by over 20%.
• ARMADA leads to a more than 4× increase in success rate and a greater than 2× decrease in human intervention ratio compared to previous human-in-the-loop learning approaches that require full-time human supervision.
• ARMADA conduces to saliently larger improvement in task progress and data efficiency using more robots in parallel, and expedites policy adaptation to novel scenarios.

🛠️ Installation

💻 Conda Environment

We test our codebase on Python 3.10. Please create an environment named armada using the following command.

conda env create -f conda_environment.yaml

If you'd like to test on real robot, execute the following command.

conda env create -f conda_environment_real.yaml

🦾 Hardware setup

Please refer to hardware setup guide for more information.

📷 Data Collection

The following example collects expert demonstrations with an image resolution of 224*224 and a 10Hz control frequency. Please feel free to tailor it to your own needs.

python record.py --output /path/to/your/output/path --resolution 224 224 --fps 10

🤖 Policy Training

The main training code can be found in armada/diffusion_policy, and the corresponding configuration files are placed under armada/config/training.

An example usage of our training recipe is in armada/train.sh.

🌐 Multi-Robot Deployment

Please refer to Multi-robot deployment guide for more information.

📝 TODO

• Release the training code and one-to-multiple shared control codebase.
• Release the code for multiple-to-multiple control.

✍️ Citation

@misc{yu2025armadaautonomousonlinefailure,
      title={ARMADA: Autonomous Online Failure Detection and Human Shared Control Empower Scalable Real-world Deployment and Adaptation}, 
      author={Wenye Yu and Jun Lv and Zixi Ying and Yang Jin and Chuan Wen and Cewu Lu},
      year={2025},
      eprint={2510.02298},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2510.02298}, 
}

🪪 License

This work is licensed under a Creative Commons Attribution-NonCommercial 4.0 International Public License.

🙏 Acknowledgement

Our code is built upon Diffusion Policy. Thanks for their open-source effort!