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
🎯 EPIC-Bench: A Perception-Centric Benchmark for Fine-Grained Embodied Visual Grounding in Vision-Language Models
📃 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.
Example visualization
For more, download the dataset to explore the full benchmark locally.
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 (or ModelScope):
| 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 |
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 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 is now available!
- [2026.5.15] 🚀 HuggingFace and ModelScope Dataset are available!
- [2026.5.15] 🚀 Project Page and Evaluation Code are released, the arXiv paper will come soon.
📋 Todo
- Evaluation code for EPIC-Bench
- The EPIC-Bench datasets
- Make the evaluation pipeline compatible with mask outputs
🏆 Leaderboard and Benchmark
Please refer to the EPIC-Bench Homepage for the full leaderboard, dataset downloads, and data examples.
📚 Citation
@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 file for details.