EPIC-Bench: A Perception-Centric Benchmark for Fine-Grained Embodied Visual Grounding in Vision-Language Models
EPIC-Bench is a fine-grained benchmark testing how vision-language models understand physical interactions in real-world environments. It systematically evaluates models across three stages: target localization, navigation, and manipulation. The benchmark reveals that current VLMs struggle with complex tasks like counting multiple targets and understanding part-whole relationships.