RoboFine-Bench: A Fine-Grained Robotic Video Understanding Benchmark

RoboFine-Bench is a benchmark for evaluating Vision-Language Models on execution-level details of robot manipulation. It contains 500 held-out robot videos and is part of the FineVLA framework for fine-grained instruction alignment. The dataset was authored by xlangai and last updated on HuggingFace in May 2026.

Use Cases

• Benchmarking VLM performance on fine-grained robot action recognition based on video data.
• Training models to understand how a robot performs a task, not just what task it is performing.
• Advancing Vision-Language-Action learning for more precise robotic control.

Strengths

• Benchmark contains 500 held-out robot videos.
• Specifically designed for evaluating fine-grained, execution-level understanding.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count and file size are unknown, which may limit suitability assessment.

Provenance

Source

HuggingFace dataset authored by xlangai.

Freshness

Last updated 2026-05-25 18:49:35.