Robo-Dopamine-GRM-Dataset

A large-scale vision-language dataset for general robotic process reward modeling.

     

Overview

Robo-Dopamine studies general process reward modeling for robotic manipulation. The core GRM setting asks a vision-language model to compare robot states under a task instruction and estimate whether an AFTER state has made progress over a BEFORE state. This dataset provides image references and instruction-following JSON annotations for training or evaluating such progress reward models.

Each sample contains:

• a task instruction,
• image paths for reference start/end anchors,
• multi-view BEFORE images,
• multi-view AFTER images,
• a conversation prompt asking for a progress score,
• a target response formatted as a score tag.

The score represents relative task progress and is formatted as:

<score>+NN%</score>
<score>-NN%</score>
<score>0%</score>

Files

The released files are organized as:

Robo-Dopamine-GRM-Dataset/
|-- images/
|   |-- dopamine.zip.part_00
|   |-- dopamine.zip.part_01
|   |-- dopamine.zip.part_02
|   |-- dopamine.zip.part_03
|   `-- dopamine.zip.part_04
|-- json_all.zip
`-- json_sel.zip

images/dopamine.zip.part_* are byte-level split parts of one ZIP archive. They are not standalone ZIP files. Concatenate them in order before unzipping.

File	Description	Size
images/dopamine.zip.part_00	image ZIP part 0	50G
images/dopamine.zip.part_01	image ZIP part 1	50G
images/dopamine.zip.part_02	image ZIP part 2	50G
images/dopamine.zip.part_03	image ZIP part 3	50G
images/dopamine.zip.part_04	image ZIP part 4	39G
json_all.zip	full JSON annotations	6.1G
json_sel.zip	balanced-sampled JSON annotations for fast model training	544M

The image archive expands to paths beginning with dopamine/, matching the image paths stored in the JSON files.

Annotation Splits

json_all.zip contains 18 JSON files and 34,601,209 samples.

json_sel.zip contains 18 JSON files and 2,997,257 samples. It is obtained by balanced sampling from json_all and is intended for quick use when training your own models.

Source subset	json_all samples	json_sel samples
agibotworld	3,160,627	316,062
agilex_newdragon	283,340	113,336
agilex_part0	982,640	196,528
agilex_part1	266,492	106,596
agilex_part2	261,224	104,489
agilex_part3	476,291	95,258
agilex_part4	345,780	69,156
agilex_task1	298,201	89,460
agilex_task2	83,000	41,500
droid_oxe	17,966,022	898,301
dual_franka_four_task	25,268	25,268
galaxea_r1lite	695,950	139,190
human_egodex	6,610,276	330,513
human_pika	573,920	57,392
libero_data	1,330,196	133,019
robocasa_data	523,042	104,608
robotwin_part1	677,948	135,589
robotwin_part2	40,992	40,992
Total	34,601,209	2,997,257

Restore The Dataset

Download the repository, then reconstruct and unzip the image archive:

cd Robo-Dopamine-GRM-Dataset

# Reconstruct the original image ZIP.
cat images/dopamine.zip.part_* > dopamine.zip

# Extract images. This creates a dopamine/ directory.
unzip dopamine.zip

# Extract annotations.
unzip json_all.zip
unzip json_sel.zip

If you want a smaller balanced subset for faster model training, unzip json_sel.zip and the image archive.

JSON Format

Each JSON file is a list of training examples. A typical example has the following structure:

{
  "id": "episode-id-bf_000010-af_000130-plus-0000179",
  "image": [
    "dopamine/.../reference_start.png",
    "dopamine/.../reference_end.png",
    "dopamine/.../before_front.png",
    "dopamine/.../before_left_wrist.png",
    "dopamine/.../before_right_wrist.png",
    "dopamine/.../after_front.png",
    "dopamine/.../after_left_wrist.png",
    "dopamine/.../after_right_wrist.png"
  ],
  "task": "pick and place",
  "conversations": [
    {
      "from": "human",
      "value": "... prompt with <image> placeholders ..."
    },
    {
      "from": "gpt",
      "value": "<score>+100.0%</score>"
    }
  ]
}

The image field usually contains 8 paths:

1. reference start front image,
2. reference end front image,
3. BEFORE front image,
4. BEFORE left wrist image,
5. BEFORE right wrist image,
6. AFTER front image,
7. AFTER left wrist image,
8. AFTER right wrist image.

For single-view sources, some camera views may map to dataset-specific camera names, but the prompt follows the same reference/BEFORE/AFTER comparison convention.

Loading Example

import json
import zipfile

zip_path = "json_sel.zip"

with zipfile.ZipFile(zip_path) as zf:
    json_files = [name for name in zf.namelist() if name.endswith(".json")]
    with zf.open(json_files[0]) as f:
        data = json.load(f)

example = data[0]
print(example["id"])
print(example["task"])
print(example["image"])
print(example["conversations"][-1]["value"])

For large-scale training, prefer streaming or chunked preprocessing instead of loading all JSON files into memory at once.

Setup

The Robo-Dopamine codebase can be installed with:

git clone https://github.com/FlagOpen/Robo-Dopamine.git
cd Robo-Dopamine

conda create -n robo-dopamine python=3.10
conda activate robo-dopamine
pip install -r requirements.txt

Related Models

• Robo-Dopamine-GRM-2.0-4B-Preview
• Robo-Dopamine-GRM-2.0-8B-Preview
• Robo-Dopamine-GRM-8B
• Robo-Dopamine-GRM-3B

Citation

If you find this dataset or project useful, please cite:

@article{tan2025robo,
  title={Robo-Dopamine: General Process Reward Modeling for High-Precision Robotic Manipulation},
  author={Tan, Huajie and Chen, Sixiang and Xu, Yijie and Wang, Zixiao and Ji, Yuheng and Chi, Cheng and Lyu, Yaoxu and Zhao, Zhongxia and Chen, Xiansheng and Co, Peterson and others},
  journal={arXiv preprint arXiv:2512.23703},
  year={2025}
}

License

This dataset card follows the Robo-Dopamine release style and uses the Apache-2.0 metadata tag. Users should also respect the licenses and usage terms of the original robot datasets represented in the image and annotation sources.