RoboFine-Bench

RoboFine-Bench is a benchmark for evaluating Vision-Language Models on fine-grained robotic manipulation understanding. It tests whether VLMs can comprehend execution-level details in robot videos, including contact regions, approach directions, motion trajectories, gripper states, and temporal ordering.

Dataset Structure

RoboFine-Bench/
├── videos/                    # Robot manipulation videos (mp4)
│   ├── BC-Z/
│   ├── BridgeDataV2/
│   ├── DROID-Robointer/
│   ├── Galaxea/
│   ├── RDT/
│   ├── RH20T-RoboInter/
│   ├── RT-1/
│   ├── RoboCoin/
│   ├── RoboMINDV1/
│   └── RoboMINDV2/
├── EvalSets.json              # Caption evaluation set (GT annotations)
├── QAEvalSets.json            # VQA evaluation set (questions + answers)
└── GT_AtomicFacts.jsonl       # Atomic facts for caption scoring

Evaluation Metrics

RoboFine-Bench evaluates VLMs through two complementary metrics:

1. VQA Evaluation

Multiple-choice questions targeting 13 fine-grained capability dimensions. Models receive video frames and answer questions about manipulation details.

2. Caption Evaluation (Atomic Fact Alignment)

Models generate dense captions for robot videos. Captions are decomposed into atomic facts and aligned against human-annotated ground truth across 13 dimensions:

• action_primitive, actor_identity, object_recognition, object_disambiguation
• contact_region, source_state_or_location, trajectory_and_orientation
• placement_specification, interaction_with_other_objects
• success_failure_retry, gripper_state, temporal_order_and_step_boundary, body_motion

Scoring measures: coverage (GT facts recalled), consistency (no contradictions), anti-hallucination (no invented facts).

Data Sources

Videos are sourced from 10 public robot manipulation datasets spanning diverse embodiments, environments, and task complexities.

Usage

from datasets import load_dataset

dataset = load_dataset("FineVLA/RoboFine-Bench")

For evaluation code, see: github.com/EricsXt/FineVLA

Citation

@article{hu2026finevla,
  title={FineVLA: Fine-Grained Instruction Alignment for Steerable Vision-Language-Action Policies},
  author={Hu, Xintong and Huang, Xuhong and Zhang, Jinyu and Yao, Yutong and Sun, Yuchong and Wang, Qiuyue and Li, Mingsheng and Liu, Yitao and Chen, Yixuan and Zheng, Yingming and Bai, Shuai and Yu, Tao},
  year={2026}
}