ruichunma's picture
Add files using upload-large-folder tool
ce86585 verified
|
Raw
History Blame Contribute Delete
7.94 kB
# Muse_Robot_Dataset_02
机器人饮料抓取与放置任务数据集,包含多模态传感器数据(RGB、声学、雷达、热成像、深度)与机械臂控制信号。
## 1. 数据概况
| 项目 | 数值 |
|------|------|
| Episode 总数 | 118 |
| 有效 Episode | 97 |
| 废弃 Episode | 21(标注 `problem.abolished = true`) |
| 任务类型 | 饮料抓取 + 放置(4 种指令) |
| 每个 Episode 帧数 | 271–483(平均 ~364) |
| RGB 分辨率 | 1280 x 720 |
## 2. 目录结构
```
Muse_Robot_Dataset_02/
├── README.md
├── thermal_data/ # 118 个 episode 数据目录
│ └── {episode_name}/
│ ├── images/ # RGB 图片 (1280x720, .jpg)
│ ├── image_timestamps.txt # 帧时间戳
│ ├── video.mp4 # [生成] RGB 视频
│ ├── mask/
│ │ └── video.mp4 # [生成] SAM3 分割 mask 视频
│ ├── depth/
│ │ └── images/ # 深度图 (1280x720, .jpg)
│ ├── acoustic/
│ │ ├── images/ # 声学频谱图 (699x809, .jpg)
│ │ └── images_masked/ # [生成] 声学 overlay 视频
│ │ ├── video.mp4 # 彩色背景
│ │ ├── video_dark.mp4 # 黑色背景
│ │ └── video_grey.mp4 # 灰度背景
│ ├── radar/
│ │ ├── images/ # 雷达图 (182x182, .jpg)
│ │ ├── raw/ # 雷达原始数据 (.npy)
│ │ └── images_masked/ # [生成] 雷达 overlay 视频
│ │ ├── video.mp4
│ │ ├── video_dark.mp4
│ │ └── video_grey.mp4
│ ├── thermal/
│ │ ├── left/
│ │ │ ├── images/ # 左热像仪 (192x256, .jpg)
│ │ │ └── raw/ # 左热像仪原始数据 (.npy)
│ │ ├── right/
│ │ │ ├── images/ # 右热像仪 (192x256, .jpg)
│ │ │ └── raw/ # 右热像仪原始数据 (.npy)
│ │ ├── images/
│ │ │ └── video.mp4 # [生成] 左右热像合并视频
│ │ └── images_masked/ # [生成] 热成像 overlay 视频
│ │ ├── video.mp4
│ │ ├── video_dark.mp4
│ │ └── video_grey.mp4
│ ├── 6d_pose/
│ │ └── right_arm.npy # 6D 位姿
│ ├── arm_command/
│ │ └── right_arm.txt # 关节指令
│ ├── arm_status/
│ │ ├── raw_right_arm.txt # 原始关节状态
│ │ └── right_arm.txt # 关节状态
│ ├── ee_command/
│ │ └── right_ee.txt # 末端执行器指令
│ └── hand_status/
│ ├── right_hand_force.txt # 手指力
│ └── right_hand_position.txt # 手指位置
└── thermal_labels/ # 标注文件
├── {episode_name}.json # 118 个标注文件
└── {episode_name}_validity.json # 118 个有效性文件 [生成]
```
> 标记 `[生成]` 的文件由 `scripts/process_dataset_02.py` 处理生成。
## 3. 传感器数据
### 图像分辨率
| 传感器 | 分辨率 | 格式 | 说明 |
|--------|--------|------|------|
| RGB | 1280 x 720 | .jpg | 主相机 |
| Depth | 1280 x 720 | .jpg | 深度图 |
| Acoustic | 699 x 809 | .jpg | 声学频谱图 |
| Radar | 182 x 182 | .jpg / .npy | 雷达图像 + 原始数据 |
| Thermal (L/R) | 192 x 256 | .jpg / .npy | 左/右热像仪 + 原始数据 |
### 帧对齐
- RGB、Acoustic、Thermal、Depth、Radar 帧数一致,帧号对齐
- 生成的 overlay 视频已做帧对齐处理:缺帧用最近邻帧填充
## 4. 标注格式
### Label 文件(`thermal_labels/{episode}.json`)
```json
{
"video_name": "2026_04_07_21_02_52_000",
"relative_path": "video/2026_04_07_21_02_52_000",
"description": "Video annotation file.",
"frame_num_total": 453,
"pre_instructions": ["Pick up the hot drink in the box.", "..."],
"problem": {
"abolished": false,
"issue": false
},
"annotations": [
{
"instruction": "Pick up the cold drink in the box.",
"start": 40,
"end": 313
},
{
"instruction": "Place the drink into the basket.",
"start": 300,
"end": 452
}
]
}
```
**字段说明**
- `video_name`: Episode 名称
- `frame_num_total`: 总帧数
- `problem.abolished`: 是否废弃(true = 数据有问题,不可用于训练)
- `problem.issue`: 是否有潜在问题
- `annotations`: 标注片段列表,每个包含 `instruction`(任务指令)、`start`/`end`(帧范围)
- `pre_instructions`: 预定义指令模板列表
### Validity 文件(`thermal_labels/{episode}_validity.json`)
由处理流水线生成,记录 SAM3 分割验证结果:
```json
// 有效 episode
{"abolished": false, "reason": "Valid: 2 instances detected"}
// 废弃 episode(标注中标记)
{"abolished": true, "reason": "Marked abolished in label"}
// 废弃 episode(SAM3 检测不足)
{"abolished": true, "reason": "SAM3 detected 1 instance(s), expected >= 2"}
```
## 5. 任务指令
共 4 种唯一指令,每个有效 episode 包含 2 个标注片段(抓取 + 放置):
| 指令 | 出现次数 |
|------|----------|
| `Pick up the hot drink in the box.` | 32 |
| `Pick up the cold drink in the box.` | 30 |
| `Pick up the room-temperature drink in the box.` | 35 |
| `Place the drink into the basket.` | 96 |
> 共 193 个标注片段,分布于 97 个有效 episode 中。
## 6. 生成的视频文件
通过 `scripts/process_dataset_02.py` 处理生成,仅对 97 个有效 episode 产出:
| 视频 | 路径 | 说明 |
|------|------|------|
| RGB 视频 | `video.mp4` | RGB 图片编码为 H.264 视频 |
| Mask 视频 | `mask/video.mp4` | SAM3 分割 mask(白色=目标区域) |
| 声学 raw | `acoustic/images/video.mp4` | 声学频谱原始视频 |
| 声学 overlay | `acoustic/images_masked/video{,_dark,_grey}.mp4` | 声学 overlay(3 种背景) |
| 雷达 raw | `radar/images/video.mp4` | 雷达原始视频 |
| 雷达 overlay | `radar/images_masked/video{,_dark,_grey}.mp4` | 雷达 overlay(3 种背景) |
| 热成像 左 raw | `thermal/left/images/video.mp4` | 左热像仪原始视频 |
| 热成像 右 raw | `thermal/right/images/video.mp4` | 右热像仪原始视频 |
| 热成像 合并 | `thermal/images/video.mp4` | 左右热像合并视频 |
| 热成像 overlay | `thermal/images_masked/video{,_dark,_grey}.mp4` | 热成像 overlay(3 种背景) |
**每个有效 episode 共 16 个 mp4 文件**,所有视频帧数与 RGB 一致。
### 处理参数
| 参数 | 值 | 说明 |
|------|-----|------|
| SAM3 Prompt | `small box` | 检测饮料盒 |
| score_threshold_detection | 0.65 | 初始检测置信度阈值 |
| mask_area_threshold | 10000 | 面积过滤:> 10000 像素的 mask 被滤除(大盒子) |
| 视频编码 | H.264, CRF 23, 30fps | — |
## 7. 与 Dataset_01 的差异
| | Dataset_01 | Dataset_02 |
|---|---|---|
| Episode 数量 | 454 | 118(97 valid + 21 abolished) |
| 顶层结构 | `{sensor}_data/` + `{sensor}_labels/` 分离 | 仅 `thermal_data/` + `thermal_labels/`,所有传感器在 episode 内部 |
| 任务类型 | towel/clothes/box/item/drink(11 种指令) | 仅 drink(4 种指令) |
| SAM3 Prompt | 按传感器分别设置 | 统一 `small box` |
| 帧对齐 | 未对齐(sensor 视频可能帧数不同) | 已对齐(缺帧用最近邻填充) |