README.md

DROID Robot Manipulation Dataset (Sampled)

数据集概述

这是从 DROID 1.0.1 数据集中采样的机器人操作失败案例子集。

• 总样本数: 2064
• 数据类型: failure
• 采样策略: balanced
• 视频格式: MP4, 60fps, 1280x720

数据集结构

hg_data/
├── videos/           # 视频文件
│   ├── 0000.mp4
│   ├── 0001.mp4
│   └── ...
├── metadata/         # 元数据文件
│   ├── 0000.json
│   ├── 0001.json
│   └── ...
├── dataset_info.json # 数据集总体信息
└── README.md         # 本文件

任务类别分布

任务类别	数量	占比
Open a drawer and take some items out	4	0.2%
Place objects in cabinet and close it	4	0.2%
Take disk off peg	4	0.2%
close book	4	0.2%
fold cloth	4	0.2%
fold the cloth	4	0.2%
open book	4	0.2%
pen in cup	4	0.2%
pour coffee beans in bowl	4	0.2%
stir with spoon	4	0.2%
wipe surface	4	0.2%
Arrange objects in a line	4	0.2%
Clump objects (cloth, rope, plush toy)	4	0.2%
Clump objects (cloth, rope, stuffed toy)	4	0.2%
Do any task, and then reset the scene.

Suggested task:

• Fo | 4 | 0.2% | | Do any task, and then reset the scene.

Suggested task:

• Ha | 4 | 0.2% | | Do any task, and then reset the scene.

Suggested task:

• Mo | 4 | 0.2% | | Do any task, and then reset the scene.

Suggested task:

• Mo | 4 | 0.2% | | Do any task, and then reset the scene.

Suggested task:

• Mo | 4 | 0.2% | | Do any task, and then reset the scene.

Suggested task:

• Op | 4 | 0.2% | | ... 其他 614 个任务 | ... | ... |

数据源分布

数据源	数量
AUTOLab	867
RPL	270
ILIAD	175
REAL	141
IPRL	139
TRI	134
WEIRD	128
IRIS	80
RAIL	56
CLVR	43
GuptaLab	30
PennPAL	1

元数据格式

每个视频对应一个 JSON 元数据文件，包含以下字段：

{
  "video_id": "0000",
  "uuid": "...",
  "lab": "数据源名称",
  "current_task": "任务描述",
  "success": false,
  "trajectory_length": 305,
  "date": "2023-10-29",
  "timestamp": "2023-10-29-20h-30m-18s",
  ...
}

使用方法

加载数据集

import json
from pathlib import Path

# 读取数据集信息
with open('dataset_info.json', 'r') as f:
    dataset_info = json.load(f)

# 遍历所有样本
for sample in dataset_info['samples']:
    video_path = sample['video_file']
    metadata_path = sample['metadata_file']

    # 读取元数据
    with open(metadata_path, 'r') as f:
        metadata = json.load(f)

    print(f"Video: {video_path}")
    print(f"Task: {metadata['current_task']}")

使用 HuggingFace Datasets

from datasets import load_dataset

dataset = load_dataset("your-username/droid-failure-sampled")

for example in dataset['train']:
    video = example['video']
    metadata = example['metadata']
    print(f"Task: {metadata['current_task']}")

引用

如果使用此数据集，请引用原始 DROID 论文：

@article{droid2024,
  title={DROID: A Large-Scale In-the-Wild Robot Manipulation Dataset},
  author={...},
  journal={...},
  year={2024}
}

许可证

与原始 DROID 数据集保持一致。

联系方式

如有问题，请联系数据集维护者。