README.md

metadata

tags:
  - robo-vqa
  - embodied-ai
  - visual-question-answering
  - video-understanding
  - physbrain-evalkit

RoboVQA 16-Frame Evaluation Dataset

This dataset package provides the 16-frame RoboVQA evaluation data used by the PhysBrainEvalKit project. It is intended for reproducible evaluation of vision-language models on the RoboVQA benchmark with the PhysBrainEvalKit eval_robovqa.py adapter.

Contents

• qa.jsonl: 1,921 evaluation examples. Each row contains the question, answer metadata, and 16 ordered frame paths.
• clips.jsonl: clip-level metadata used to identify the source video clips.
• metadata.json: dataset construction and sample-count metadata.
• frames.tar: 21,360 JPEG frames from 1,335 unique clips. The archive stores the files under the frames/ prefix.

The evaluation protocol expects exactly 16 time-ordered frames for every QA example. Frame paths in qa.jsonl are resolved relative to this dataset directory or to members in frames.tar.

Use with PhysBrainEvalKit

Set the dataset root to the directory containing qa.jsonl and frames.tar:

export ROBOVQA_DATA_ROOT=/path/to/RoboVQA-16frames

Then run the RoboVQA benchmark from the PhysBrainEvalKit repository:

python eval_robovqa.py \
  --model_path /path/to/qwen3_vl_model \
  --model_name qwen3-vl \
  --backbone qwen3 \
  --backend hf \
  --data_root "$ROBOVQA_DATA_ROOT" \
  --expected_num_frames 16

The PhysBrainEvalKit loader reads images from frames.tar on demand. Manual extraction is not required. Keeping the archive intact avoids creating thousands of separate image files.

File format

qa.jsonl is newline-delimited JSON. The loader requires a non-empty frame_paths list in every row and validates that it contains 16 entries. Relative frame paths must match the corresponding members in frames.tar.

Provenance and licensing

This package is a prepared evaluation layout for PhysBrainEvalKit. RoboVQA content and any underlying media remain subject to the terms of the original RoboVQA dataset and its source datasets. Review the upstream dataset license and citation requirements before redistribution or publication.