FailSafe: Reasoning and Recovery from Failures in Vision-Language-Action Models

Introduction

FailSafe is a failure generation and recovery system that automatically produces diverse failure cases paired with executable recovery actions. It can be seamlessly applied to any manipulation task in any simulator, enabling scalable creation of failure–action data.

Prerequisite

Install dependencies in virtual environment:

# Create a conda environment
conda env create -f environment.yml -n failsafe

• Please refer to Maniskill to download the required packages.

Get Started

1. Failure Generation

python examples/ex_failgen_data_collection.py --headless --save-video

2. Action Collection and Systematic Verification

• Change the TBD in the sanity_check.py into your data path

cd data
python sanity_check.py

3. Ground-truth Collection

• Change the TBD in the gt_collection.py into your data path

python gt_collection.py

Acknowledgments

We gratefully acknowledge the open-source projects that our work builds upon.

• Maniskill
• AHA

✉️ Feel free to email me ([email protected]) or raise the issue if you have any questions about our work.

Citation

If you find our work useful in your research, please cite it as follows:

@misc{lin2025failsafe,
      title={FailSafe: Reasoning and Recovery from Failures in Vision-Language-Action Models}, 
      author={Zijun Lin and Jiafei Duan and Haoquan Fang and Dieter Fox and Ranjay Krishna and Cheston Tan and Bihan Wen},
      year={2025},
      eprint={2510.01642},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2510.01642}, 
}