RoboStressBench/RoboStressBench-Dataset

RoboStressBench

RoboStressBench is a robotics-oriented vision-language benchmark for evaluating model robustness under visual stress conditions. It contains two task formats: visual question answering and localization tasks.

Task Formats

Visual Question Answering

• Reasoning: interpret the scene under stress.
• Planning: decide the next embodied action.
• Spatial: answer relative-position and layout questions.
• State Understanding: judge task completion or object state.

Localization Tasks

• Placement Grounding: point to free space for placing an object.
• Target Grounding: localize the target object with a bounding box or point.

Dataset Structure

manifest.jsonl
metadata.jsonl
<source_id>/<subcategory_id>/records/<sample_id>.json
<source_id>/<subcategory_id>/media/rgb/<image>
<source_id>/<subcategory_id>/media/gt_mask/<mask>

manifest.jsonl is the canonical index used by the evaluation code. metadata.jsonl is a flattened convenience file for browsing and external analysis.

Each record contains:

• instruction: the question or grounding instruction.
• rgb: relative path to the input image.
• stress: first- and second-level stress labels.
• gt: ground truth, with one of mcq, bbox, or mask.

Bounding boxes use xyxy coordinates in permille space [0, 1000]. Mask/placement tasks expect a predicted point (x, y) in the same permille coordinate space.

License and Terms

This dataset is released under the RoboStressBench Research-Only Non-Commercial Dataset License. See LICENSE.

RoboStressBench is constructed from existing public benchmarks, Pexels-sourced real-world images, and controlled stress synthesis. We retain the license and usage terms of each original data source. Our annotations, metadata, and benchmark construction code may be released under our chosen research license, while images and derived visual assets remain subject to the licenses or terms of their corresponding source data.

In particular:

• Existing public benchmark sources include datasets released under CC BY 4.0 or Apache 2.0 terms, as well as RoboRefit samples distributed without an explicit dataset license and used here for non-commercial academic research only.
• Pexels-sourced images remain subject to the Pexels License and may not be redistributed or sold as standalone stock photos, wallpaper assets, or similar image collections.
• Controlled stress synthesis samples inherit the license and usage constraints of their underlying source datasets. Synthetic samples based on proprietary in-house data are released for research-only, non-commercial use.

Users are responsible for complying with all applicable source licenses and terms.

Citation

@misc{wu2026robostressbenchbenchmarkingvlmrobustness,
      title={RoboStressBench: Benchmarking VLM Robustness to Physical Visual Stress in Embodied Scenes}, 
      author={Leyi Wu and Yifan Zhao and Jinjie Zhang and Suzeyu Chen and Wosong Chen and Zhifei Chen and Tianshuo Xu and Qingchun He and Hongxin Hu and Haojian Huang and Yangkai Wei and Wenqian Li and Yinchuan Li and Ying-Cong Chen},
      year={2026},
      eprint={2606.00828},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2606.00828}, 
}