Awesome-VLA-Safety

KO | EN

gitlite — search

#python #java #python3 #arduino #golang #machine-learning #rust #html #flask #javascript #seismology #nodejs

Awesome-VLA-Safety

★ 11 Open GitHub ↗

A curated paper list for Vision-Language-Action (VLA) safety, accompanying our survey "Safety of Vision-Language-Action Models: A Survey from Lifecycle Perspectives".

// repository documentation

Was this content helpful?

★ 0 (0 ratings)

# Awesome VLA Safety Papers: Vision-Language-Action Model Safety Paper List

[![Awesome](https://cdn.rawgit.com/sindresorhus/awesome/d7305f38d29fed78fa85652e3a63e154dd8e8829/media/badge.svg)](https://github.com/sindresorhus/awesome)
<img src="https://img.shields.io/badge/Contributions-Welcome-278ea5" alt="Contrib"/>

---
A curated and continuously updated paper repository for Vision-Language-Action (VLA) safety, covering VLA attacks, defenses, robustness, safety evaluations from data safety, training safety, and deployment safety perspectives.

This repository provides a **comprehensive and structured resource** for Vision-Language-Action (VLA) safety, based on our survey:   
📄 **[Safety of Vision-Language-Action Models: A Survey from Lifecycle Perspectives](https://www.authorea.com/doi/full/10.22541/au.177524426.60806944/v1)**

---

**Vision-Language-Action (VLA) models** enable embodied agents to perceive, reason, and act in real-world environments through unified end-to-end multimodal policies.  
As VLA models are increasingly deployed in real-world environments, **ensuring their safety has become a critical concern**, as failures or malicious behaviors may lead to severe physical harm and broader societal consequences.

**The Scope of Safety in This Survey.**
In this repository, the safety in VLA systems is defined as the ability to learn a reliable mapping from multimodal observations and instructions to effective actions, avoiding the unsafe behaviors, harmful consequences, and degraded system reliability.

# 📢 Latest Updates

[21/04/2026] Update the taxonomy and add some new papers.  
[29/03/2026] Repo initialized.

---

# 📚 Contents

- [Awesome VLA Safety Papers](#awesome-vla-safety-papers)
- [📢 Latest Updates](#-latest-updates)
- [📚 Contents](#-contents)
- [🧠 Taxonomy](#-taxonomy)
- [📄 Relevant Survey](#-relevant-survey)
- [🗂 Data Safety](#-data-safety)
  - [Training Data Poisoning](#training-data-poisoning)
  - [Data Scarcity](#data-scarcity)
  - [Data Quality](#data-quality)
- [📈 Training Safety](#-training-safety)
  - [Pretraining-Inherited \& Finetuning Risks](#pretraining-inherited--finetuning-risks)
  - [Inappropriate Objective Design](#inappropriate-objective-design)
  - [Generalization Gap](#generalization-gap)
- [🚀 Deployment Safety](#-deployment-safety)
  - [Deployment-Time Attacks](#deployment-time-attacks)
  - [External Uncertainty](#external-uncertainty)
  - [Inappropriate Use](#inappropriate-use)
- [📬 Citation](#-citation)
- [🤝 Contribution](#-contribution)

---

# 🧠 Taxonomy

We organize VLA safety from a **lifecycle perspective**:

- **Data Preparation → Data Safety**
- **Model Training → Training Safety**
- **System Deployment → Deployment Safety**

<!-- Each stage is further divided into:

- 🔴 **Adversarial Scenarios**
- 🟢 **Non-adversarial Scenarios**   -->

The following figure shows lifecycle-oriented taxonomy of our VLA safety survey.
<p align="center">
  <figure>
    <img src="./taxonomy.jpg" alt="VLA Safety Taxonomy" width="750">
    
  </figure>
</p>

---

# 📄 Relevant Survey

---

# 🗂 Data Safety

---

# 📈 Training Safety

## Pretraining-Inherited & Finetuning Risks  
Note: most data poisoning attacks need to revise the training process to strengthen their adversarial effects. Therefore, this part may have large overlaps with [Training Data Poisoning](#training-data-poisoning).

---

# 🚀 Deployment Safety

# 📬 Citation

If you feel the repository and survey are helpful, please cite us ~ 
```bibtex
@article{yuan2026safety,
  title={Safety of Vision-Language-Action Models: A Survey from Lifecycle Perspectives},
  author={Yuan, Wei and Liu, Fengwen and Wei, Ruize and Wang, Zongwei and Gao, Yuan and Huang, Hu and Fang, Binxing and Liao, Qing},
  journal={Authorea Preprints},
  year={2026},
  publisher={Authorea}
}
```

# 🤝 Contribution

We welcome contributions to improve this repository!

If you have relevant papers, feel free to submit a pull request:

---

Awesome-VLA-Safety

Explore Similar Repositories

Was this content helpful?

Recent Feedback

Download README