---
tags:
- vlabench
- embodied-ai
- visual-question-answering
- robotics
- physbrain-evalkit
---

# VLABench VLM Evaluation Dataset

This dataset is the VLM evaluation split of [VLABench](https://vlabench.github.io/), prepared for reproducible VLABench evaluation with [PhysBrainEvalKit](https://github.com/DeepCybo-PhysAI/PhysBrainEvalKit).

## Source

- Project page: <https://vlabench.github.io/>
- Paper: <https://arxiv.org/abs/2412.18194>
- Official code: <https://github.com/OpenMOSS/VLABench>

## Directory layout

The dataset is organized by evaluation dimension and subtask:

```text
vlm_evaluation_v1.0/
├── CommenSence/
├── Complex/
├── M&T/
├── PhysicsLaw/
├── Semantic/
└── Spatial/
```

Each subtask contains evaluation episodes, typically `example0` through `example99`:

```text
<dimension>/<subtask>/example0/
├── input/
│   ├── input.png
│   └── input_mask.png
├── instruction/
│   └── input_instruction.txt
├── output/
│   └── operation_sequence.json
└── env_config/
```

The PhysBrainEvalKit evaluator requires `input/input.png`, `input/input_mask.png`, `instruction/input_instruction.txt`, and `output/operation_sequence.json`. The optional `env_config/` directory contains episode configuration files for users who want to reproduce the underlying simulation environment; it is not required for VLM scoring.

## Use with PhysBrainEvalKit

Download or clone this dataset directory, then set its local path:

```bash
export VLABENCH_DATASET_PATH=/path/to/vlm_evaluation_v1.0
```

Run the VLABench benchmark from the PhysBrainEvalKit repository:

```bash
python eval_vlabench.py \
  --model_path /path/to/qwen3_vl_model \
  --model_name qwen3-vl \
  --backbone qwen3 \
  --backend hf \
  --dataset_path "$VLABENCH_DATASET_PATH"
```

The standard launcher uses the same path automatically when `VLABENCH_DATASET_PATH` is set:

```bash
bash scripts/eval_qwen3vl.sh \
  --model-path /path/to/qwen3_vl_model \
  --model-name qwen3-vl \
  --output-base /path/to/results \
  --gpus 0,1 \
  --only VLABench
```

## Citation

If you use VLABench, please cite the original paper:

```bibtex
@misc{zhang2024vlabench,
  title={VLABench: A Large-Scale Benchmark for Language-Conditioned Robotics Manipulation with Long-Horizon Reasoning Tasks},
  author={Shiduo Zhang and Zhe Xu and Peiju Liu and Xiaopeng Yu and Yuan Li and Qinghui Gao and Zhaoye Fei and Zhangyue Yin and Zuxuan Wu and Yu-Gang Jiang and Xipeng Qiu},
  year={2024},
  eprint={2412.18194},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2412.18194}
}
```

## License and redistribution

VLABench data and media remain subject to the license and redistribution terms of the original project. Review the upstream repository before publishing a mirror or redistributing the files.