# RoboPulse

RoboPulse is a benchmark introduced in [PRM-as-a-Judge: A Dense Evaluation Paradigm for Fine-Grained Robotic Auditing](https://arxiv.org/abs/2603.21669) for testing whether a vision-language judge can detect fine-grained relative progress in physical manipulation.

This Hugging Face release contains the hard `1800`-example subset. Each example asks the judge to compare a `BEFORE` state and an `AFTER` state under the same task, while using task-start and task-end reference frames as anchors for the full task scope.

## Overview

The figure below illustrates the multi-view comparison setup in RoboPulse.

![RoboPulse visual overview](./robopulse_vis.png)

## Files

- `RoboPulse.json`: benchmark annotations with release-relative image paths
- `images.zip`: zipped image assets
- `README.md`: dataset overview and field definitions
- `robopulse_vis.png`: multi-view comparison illustration
- `robopulse_stat.png`: dataset coverage statistics
- `results.png`: benchmark results table from the paper

If you want the image paths in `RoboPulse.json` to resolve locally, extract `images.zip` in the same folder so that the `images/` directory sits next to `RoboPulse.json`.

## Dataset Summary

- Number of samples: `1800`
- Image references in JSON: `14400`
- Unique image files: `13059`
- Total source image size: `345.72 MB`
- Archive size: `340.97 MB`
- Source datasets: `9`
- Hop magnitude bins: `small`, `medium`, `large`

Source datasets in this release:

- `agibotworld`: `200` samples
- `agilex_newdragon`: `200` samples
- `droid_oxe`: `200` samples
- `galaxea_r1lite`: `200` samples
- `human_egodex`: `200` samples
- `human_pika`: `200` samples
- `libero_data`: `200` samples
- `robocasa_data`: `200` samples
- `robotwin2_agilex_part1`: `200` samples

The figure below summarizes the coverage of RoboPulse across data sources and task semantics.

![RoboPulse statistics](./robopulse_stat.png)

## Results

The figure below shows the main pairwise progress-judgment results reported for RoboPulse.

![RoboPulse results](./results.png)

## Data Format

Each item in `RoboPulse.json` is a dictionary with the following fields:

- `id`: unique sample identifier
- `task`: task instruction for the sample
- `image_dataset`: source dataset name
- `image`: a list of `8` image paths, all relative to this release folder
- `conversations`: question-answer style supervision for the judge
- `hop_value`: signed Hop value used to construct the sample pair
- `hop_absolute_value`: absolute value of `hop_value`
- `hop_category`: categorical metadata derived from `hop_value`

### Image Ordering

`image[0]` to `image[7]` always follow the same order:

1. `image[0]`: reference start frame for the task
2. `image[1]`: reference end frame for the completed task
3. `image[2]`: front view of the `BEFORE` state
4. `image[3]`: left wrist view of the `BEFORE` state
5. `image[4]`: right wrist view of the `BEFORE` state
6. `image[5]`: front view of the `AFTER` state
7. `image[6]`: left wrist view of the `AFTER` state
8. `image[7]`: right wrist view of the `AFTER` state

In other words, the benchmark compares a `BEFORE` triplet against an `AFTER` triplet, with start and end reference frames provided as conceptual anchors.

### Conversations

`conversations` stores the judge prompt and the target answer:

- `conversations[0]`: the evaluation question given to the judge model
- `conversations[1]`: the expected answer, such as `<score>+1</score>` for progress and `<score>-1</score>` for regression

### Hop Fields

The Hop-based fields describe the relative progress signal used to build RoboPulse. For the detailed formulation, please refer to Appendix F of the paper:

- Paper: [PRM-as-a-Judge](https://arxiv.org/abs/2603.21669)
- PDF: [https://arxiv.org/pdf/2603.21669](https://arxiv.org/pdf/2603.21669)

Field meanings:

- `hop_value`: signed relative progress change between the two compared states. Positive values indicate forward progress toward the task goal, while negative values indicate regression away from the goal.
- `hop_absolute_value`: magnitude of the progress change, ignoring direction.
- `hop_category`: a dictionary with three subfields: `absolute_category`, `direction`, and `combined_category`
- `absolute_category`: magnitude bucket of the Hop value, one of `small`, `medium`, or `large`
- `direction`: direction bucket, either `progress` (forward) or `regression` (backward)
- `combined_category`: combination of the two, such as `progress_small`, `progress_medium`, `progress_large`, `regression_small`, `regression_medium`, or `regression_large`

## Directory Layout

After extracting `images.zip`, the folder should look like this:

```text
hf_RoboPulse/
├── RoboPulse.json
├── images.zip
├── README.md
└── images/
    └── <dataset_name>/
        └── ...
```

## Usage Notes

- Upload the whole folder to your Hugging Face dataset repository.
- If you want image paths in the JSON to be directly readable from the repo, extract `images.zip` before or after uploading so that `images/` exists alongside `RoboPulse.json`.
- The release preserves the original benchmark annotations and only rewrites image paths to release-relative paths under `images/`.

## Related Links

- Project repository: [PRM-as-a-Judge](https://github.com/Yuheng2000/PRM-as-a-Judge)
- Paper: [PRM-as-a-Judge: A Dense Evaluation Paradigm for Fine-Grained Robotic Auditing](https://arxiv.org/abs/2603.21669)

## Citation

If this project, leaderboard, or evaluation pipeline helps your work, please cite:

```bibtex
@article{ji2026prmjudge,
  title = {PRM-as-a-Judge: A Dense Evaluation Paradigm for Fine-Grained Robotic Auditing},
  author = {Ji, Yuheng and Liu, Yuyang and Tan, Huajie and Huang, Xuchuan and Huang, Fanding and Xu, Yijie and Chi, Cheng and Zhao, Yuting and Lyu, Huaihai and Co, Peterson and others},
  journal = {arXiv preprint arXiv:2603.21669},
  year = {2026}
}
```