gabrieldavin23/robust-autonomous-driving

RADBench: Evaluating Foundation VLMs for Corner Case Description on A Comprehensive Dataset

Gabriel Manalu, An Vo, Do Minh Duc, Nhat-Quang Tau, Nguyen Dinh Son, Hoang Minh Son, Hyeontaek Hwang, Jaesung Lee, Hyojin Choi, Tran Tuan Phong, Daeyoung Kim

School of Computing, KAIST, Daejeon, Korea

2026 IEEE Intelligent Vehicles Symposium (IV), Oral

---

RADBench contains 522 corner-case driving scenarios (392 real dashcam clips and 130 synthetic clips) across 75 event categories. Each scenario is given to a VLM as a sequence of uniformly sampled frames with the prompt "Describe this sequence of frames in detail.", and the response is judged against a human-written ground-truth description.

Files

All data is in the dataset/ folder:

File	Contents
dataset/radbench.json	2,088 entries: 522 scenarios × 4 frame counts (4, 8, 16, 32)
dataset/MMAU.zip	286 real clips from MM-AU
dataset/UnusualData.zip	106 real clips from public dashcam footage
dataset/VEO3.zip	71 synthetic clips generated with Veo 3 from NHTSA incident reports
dataset/GTA.zip	59 synthetic clips rendered in GTA V

The main benchmark uses 16 frames. The 4, 8 and 32-frame versions are for the frame-count ablation.

Download

hf download gabrieldavin23/robust-autonomous-driving --repo-type dataset --include "dataset/*" --local-dir .
cd dataset && unzip -q '*.zip' && cd ..

This produces the layout the evaluation code expects:

dataset/
├── radbench.json
├── MMAU/sampled_frames_{4,8,16,32}/<scenario>/images/*.jpg
├── UnusualData/...
├── VEO3/...
└── GTA/...

Entry format

{
  "ID": "MMAU_014434_n_frames_16",
  "image_folder": "dataset/MMAU/sampled_frames_16/014434/images",
  "n_frames": 16,
  "image_filenames": ["000001.jpg", "000004.jpg", "..."],
  "prompt": "Describe this sequence of frames in detail.",
  "ground_truth": "...",
  "category": "ego-car hits a crossing pedestrian"
}

image_folder is relative to the directory that contains dataset/. category is empty for the synthetic VEO3 and GTA clips.

License

Several source clips are derived from MM-AU. Please consider their license and usage requirements.

Citation

@inproceedings{manalu2026radbench,
  title={RADBench: Evaluating Foundation VLMs for Corner Case Description on A Comprehensive Dataset},
  author={Manalu, Gabriel and Vo, An and Tau, Nhat-Quang and Son, Nguyen Dinh and Son, Hoang Minh and Hwang, Hyeontaek and Lee, Jaesung and Choi, Hyojin and Phong, Tran Tuan and Kim, Daeyoung and others},
  booktitle={2026 IEEE Intelligent Vehicles Symposium (IV)},
  pages={1869--1876},
  year={2026},
  organization={IEEE}
}