EPIC-Bench: A Fine-Grained Embodied Visual Grounding Benchmark for VLMs

EPIC-Bench is a Mask-Grounding-based benchmark designed to evaluate Vision-Language Models' visual perception in embodied scenarios. The benchmark covers 3 high-level categories and 23 task types, following a realistic embodied workflow. It was created by author rxc205 and last updated on 2026-05-18.

Use Cases

• Benchmarking object localization accuracy in complex scenes based on the described TargetLocalization task.
• Evaluating a VLM's ability to follow natural-language instructions in embodied scenarios based on the described workflow.
• Assessing fine-grained visual perception capabilities across 23 different task types as mentioned in the description.

Strengths

• Covers 3 high-level categories and 23 distinct task types, suggesting a structured evaluation framework.
• Designed to follow a realistic embodied workflow, likely increasing ecological validity for real-world applications.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment for large-scale training.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

huggingface

Freshness

Last updated 2026-05-18 16:26:28; freshness should be verified.