🎯 EPIC-Bench: A Perception-Centric Benchmark for Fine-Grained Embodied Visual Grounding in Vision-Language Models

📃 Overview

This repo contains the official evaluation code and dataset for the paper "EPIC-Bench: A Perception-Centric Benchmark for Fine-Grained Embodied Visual Grounding in Vision-Language Models"

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.

The goal is to measure whether models can reliably perceive the critical Visual information required throughout the Embodied Process.

✨ Highlights

• Embodied-Scenario evaluation of VLM Visual Perception capability.
• Focus on Visual Grounding / Perception without language shortcut exploitation.
• Diverse and Fine-Grained task design.

📰 News

• [2026.5.15] 🚀 HuggingFace and ModelScope Dataset are available!
• [2026.5.15] 🚀 We released the ArXiv paper.

📋 Todo

• Evaluation code for EPIC-Bench
• Make the evaluation pipeline compatible with mask outputs

🏆 Leaderboard and Benchmark

Please refer to the EPIC-Bench Homepage for:

• Leaderboard
• Full dataset downloads
• EPIC-Bench data examples

📚 Citation

@article{EPIC-Bench,
  title={EPIC-Bench: A Perception-Centric Benchmark for Fine-Grained Embodied Visual Grounding in Vision-Language Models},
  author={XXX, XXX, XXX},
  journal={},
  year={2026}
}

📜 License

Please add an explicit LICENSE file before open-sourcing. If EPIC-Bench annotations or images have redistribution constraints, publish them separately (e.g., Hugging Face / ModelScope) and keep this repo code-only + small examples.

🙏 Acknowledgements

• ms-swift for open-source VLM inference: ms-swift
• lmms-eval for API/closed-source evaluation: lmms-eval