VisionFault-920K: A Large-Scale Fault Injection Dataset for Robotic Vision Systems

A new version is available: https://zenodo.org/records/18695332 Dataset Description: A large-scale fault-augmented dataset contains 920,935 images derived from real robotic camera recordings. To simulate real-world edge cases, original frames were transformed via Stable Diffusion (img2img) across a wide range of fault scenarios. Technical Specifications • Total Images: 920,935 • Primary Tasks: Lane Following and Obstacle Detection. • LLM Engine: Fault scenarios designed using GPT-OSS 120B. • Synthesis Engine: Perturbations synthesized using Stable Diffusion 2.1 Base. •Categories: Features diverse fault types, such as Camera Failures, Motion Blur, Extreme Weather (Ice, Rain, Fog), Low Light (Tunnel, Night, Backlight), Lens Distortions, etc. Usage If "Download All" is not working, please download the files one by one. To merge the parts after downloading: For Linux/Mac: cat part* > dataset.zip For Windows (PowerShell - Recommended): Get-Content part* -ReadCount 0 -Encoding Byte | Set-Content dataset.zip -Encoding Byte For Windows (CMD): copy /b part* dataset.zip if not working : cmd /c copy /b part* dataset.zip License & Disclaimer • License: Creative Commons Attribution 4.0 International (CC BY 4.0). • Disclaimer: This dataset is LLM+LDM generated and provided "as-is" for robotic vision system testing and research.