---
license: apache-2.0
pretty_name: EPIC-Bench
task_categories:
  - visual-question-answering
  - object-detection
language:
  - en
Modalities:
  - Image
  - Text
tags:
  - embodied-perception
  - mask-grounding
  - vision-language-models
size_categories:
  - 1K<n<10K
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*.parquet
dataset_info:
  features:
    - name: image
      dtype: image
    - name: reference_object_image
      dtype: image
    - name: ground_truth_mask
      dtype: image
    - name: category
      dtype: string
    - name: task_type
      dtype: string
    - name: text_label
      dtype: string
    - name: description_en
      dtype: string
    - name: description_cn
      dtype: string
  splits:
    - name: train
      num_examples: 6661
---



<div align="center">
  
# 🎯 EPIC-Bench: A Perception-Centric Benchmark for Fine-Grained Embodied Visual Grounding in Vision-Language Models

[![arXiv](https://img.shields.io/badge/arXiv-2605.17070-b31b1b.svg)](https://arxiv.org/abs/2605.17070)
[![Project Page](https://img.shields.io/badge/Project-Page-blue)](https://epic-bench.github.io/EPIC-Bench/)
[![Dataset](https://img.shields.io/badge/🤗-Dataset-yellow)](https://huggingface.co/datasets/rxc205/EPIC-Bench)
[![Evaluation Toolkit](https://img.shields.io/badge/⚙️-Evaluation_Toolkit-6366f1.svg)](https://github.com/rxc205/EPIC-Bench-Eval#-epic-bench-evaluation-toolkit)
[![License](https://img.shields.io/badge/License-Apache_2.0-green.svg)](https://github.com/rxc205/EPIC-Bench-Eval/blob/main/LICENSE)
</div>

## 📃 Overview


📚 **EPIC-Bench** is a **Mask-Grounding-based** benchmark designed to evaluate a VLM’s **Visual Perception** capability in **Embodied Scenarios**. EPIC-Bench covers **3 High-Level Categories** and **23 Task Types**, following the realistic **Embodied Workflow**:

- 🎯 **TargetLocalization**: **Pinpoint** the right object in the scene from a natural-language instruction.
- 🧭 **Navigation**: **Approach** the target step by step by reading key visual cues along the way.
- 🤲 **Manipulation**: **Operate** on the target through fine-grained, action-oriented **Grounded Perception**.

<p align="center">
  <img src="https://epic-bench.github.io/EPIC-Bench/img/20260302-192636.png" alt="EPIC-Bench teaser" width="100%"/>
</p>


## Example visualization

<p align="center">
  <img src="https://raw.githubusercontent.com/rxc205/EPIC-Bench-Eval/refs/heads/main/images/bmk_cases.png" alt="EPIC-Bench bmk_cases" width="100%"/>
</p>
<p align="center">
  <em>For more, <a href="https://huggingface.co/datasets/rxc205/EPIC-Bench">download the dataset</a> to explore the full benchmark locally.</em>
</p>


## Download

> **Note:** The files in the `data/` directory are included solely to enable previewing examples in the Hugging Face Dataset Viewer. They are not part of the recommended usage pipeline, and users may safely ignore them when using EPIC-Bench.


EPIC-Bench contains ~35k small files across three task categories. To work around HuggingFace's per-file rate limit and to keep download speed reasonable, the annotations are distributed as **three task-level tarballs** instead of raw folders.

Download the tarballs from [HuggingFace](https://huggingface.co/datasets/rxc205/EPIC-Bench) (or [ModelScope](https://www.modelscope.cn/datasets/macarich/EPIC-Bench)):

| Archive | Size | # Files | Contents |
|---------|------|---------|----------|
| `Manipulation.tar.gz`       | 1.70 GB | 7,061   | AffordanceRegion / ContactRelationship / PlacementRegion |
| `Navigation.tar.gz`         | 2.71 GB | 10,868  | FeasiblePath / GroundDetection / VisualMatching |
| `TargetLocalization.tar.gz` | 3.34 GB | 17,665  | BasicAttributes / EmbodiedCompositionalAttributes / SpatialRelatedAttributes |

```bash
mkdir -p dataset/annotation/EPIC_Bench
cd dataset/annotation/EPIC_Bench

# Option 1: huggingface-cli
hf download rxc205/EPIC-Bench \
    --repo-type dataset --local-dir .

# Option 2: modelscope
modelscope download \
  --dataset macarich/EPIC-Bench \
  --local_dir .

# Extract all three (preserves the original folder layout)
for f in Manipulation.tar.gz Navigation.tar.gz TargetLocalization.tar.gz; do
    tar -xzf "$f" && rm "$f"
done
```

## 🧰 EPIC-Bench Evaluation Toolkit

We provide [evaluation code](https://github.com/rxc205/EPIC-Bench-Eval#-epic-bench-evaluation-toolkit) for both open-source VLMs via **ms-swift** and API-based VLMs via **lmms-eval**.  
For details, please refer to our GitHub repository.



## 📰 News

- [2026.5.19] 🎉 Our [arXiv paper](https://arxiv.org/abs/2605.17070) is now available!
- [2026.5.15] 🚀 [HuggingFace](https://huggingface.co/datasets/rxc205/EPIC-Bench) and [ModelScope](https://www.modelscope.cn/datasets/macarich/EPIC-Bench) Dataset are available!
- [2026.5.15] 🚀 [Project Page](https://epic-bench.github.io/EPIC-Bench/) and [Evaluation Code](https://github.com/rxc205/EPIC-Bench-Eval) are released, the arXiv paper will come soon.

## 📋 Todo

- [x] Evaluation code for EPIC-Bench
- [x] The EPIC-Bench datasets
- [ ] Make the evaluation pipeline compatible with mask outputs


## 🏆 Leaderboard and Benchmark

Please refer to the [EPIC-Bench Homepage](https://epic-bench.github.io/EPIC-Bench/) for the full leaderboard, dataset downloads, and data examples.

## 📚 Citation

```BibTeX
@misc{shan2026epicbenchperceptioncentricbenchmarkfinegrained,
      title={EPIC-Bench: A Perception-Centric Benchmark for Fine-Grained Embodied Visual Grounding in Vision-Language Models}, 
      author={Haozhe Shan and Xiancong Ren and Han Dong and Haoyuan Shi and Yingji Zhang and Jiayu Hu and Yi Zhang and Yong Dai and Bin Shen and Lizhen Qu and Zenglin Xu and Xiaozhu Ju},
      year={2026},
      eprint={2605.17070},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2605.17070}, 
}
```

## 📜 License


This project is licensed under the Apache License 2.0 - see the [LICENSE](https://github.com/rxc205/EPIC-Bench-Eval/blob/main/LICENSE) file for details.

## 🙏 Acknowledgements

- **ms-swift** for open-source VLM inference: [ms-swift](https://swift.readthedocs.io/zh-cn/latest/)
- **lmms-eval** for API/closed-source evaluation: [lmms-eval](https://github.com/EvolvingLMMs-Lab/lmms-eval)