You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Ragtel PiperX 双臂 1 kg 装箱数据集

本数据集由 Ragtel 双臂遥操作系统采集并转换为 LeRobot v3.0 格式,用于双臂 VLA、模仿学习和离线数据审核。当前版本包含 12 个 episode、51,185 帧,约 28.44 分钟;机器人信号和三路视频均输出为 30 Hz。

数据集概览

项目
任务 1KG装箱
机器人 双 PiperX,左右臂均为真实数据
Episode 12
总帧数 51,185
总时长 1,706.17 s(约 28.44 min)
采样率 30 Hz
State / Action float32[14]
视频 3 路真实 RGB,1280×720,H.264,30 FPS,无音频
LeRobot 版本 v3.0
Split train: 0:12

快速加载

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("TimerChen/FactoryDataset")
sample = dataset[0]

print(sample["observation.state"].shape)       # torch.Size([14])
print(sample["action"].shape)                  # torch.Size([14])
print(sample["observation.images.top"].shape)  # torch.Size([3, 720, 1280])

目录与文件规范

FactoryDataset/
├── data/chunk-000/file-000.parquet
├── videos/
│   ├── observation.images.left_wrist/chunk-000/file-000.mp4 ... file-011.mp4
│   ├── observation.images.top/chunk-000/file-000.mp4 ... file-011.mp4
│   └── observation.images.right_wrist/chunk-000/file-000.mp4 ... file-011.mp4
└── meta/
    ├── info.json
    ├── stats.json
    ├── tasks.parquet
    ├── episodes/chunk-000/file-000.parquet
    ├── ragtel/alignment.v1.json
    ├── ragtel/batch_conversion.v1.json
    └── ragtel_calibration/
        ├── manifest.json
        ├── export_intrinsics.json
        ├── robot_description.urdf
        ├── left_wrist/
        ├── top/
        └── right_wrist/

每个 episode 对应每个相机目录中的一个 MP4,文件编号与 episode_index 一致。Parquet 中的 frame_index 在每个 episode 内从 0 开始,timestamp = frame_index / 30,时间单位为秒。

字段定义

Key 类型/形状 含义
observation.state float32[14] 左右 PiperX 的实测关节位置与夹爪位置
action float32[14] 实际发送给左右 follower 的关节与夹爪目标
observation.images.left_wrist video [720,1280,3] 左腕相机,raw slot 0、采集设备 /dev/video4
observation.images.top video [720,1280,3] 顶部固定相机,raw slot 1、采集设备 /dev/video2
observation.images.right_wrist video [720,1280,3] 右腕相机,raw slot 2、采集设备 /dev/video6
timestamp float32 episode 内相对时间,单位秒
frame_index int64 episode 内帧号,从 0 开始
episode_index int64 episode 编号,范围 0–11
index int64 全数据集全局帧索引
task_index int64 任务索引;当前仅有任务 0:1KG装箱

14 维 State / Action 规范

两个向量采用完全相同的维度顺序:

[left_j1, left_j2, left_j3, left_j4, left_j5, left_j6, left_gripper,
 right_j1, right_j2, right_j3, right_j4, right_j5, right_j6, right_gripper]
  • 关节维度单位为 rad
  • 夹爪维度单位为 mm,约 0 表示闭合、90 表示张开。
  • observation.state 是机器人反馈;action 是映射后实际发送给 follower 的目标,不是下一帧 state。
  • 以下范围是当前版本 51,185 帧的实际观测 min/max,仅用于数据检查和归一化参考,不代表 PiperX 机械限位或安全控制范围。
Dim observation.state 含义 State min State max action 含义 Action min Action max
0 左 J1 位置 (rad) -1.19068 0.76754 左 J1 目标 (rad) -1.20878 0.77006
1 左 J2 位置 (rad) -0.03049 2.41850 左 J2 目标 (rad) -0.01396 2.43090
2 左 J3 位置 (rad) -2.00814 0.03383 左 J3 目标 (rad) -2.00184 0.02301
3 左 J4 位置 (rad) -1.41371 1.43562 左 J4 目标 (rad) -1.32683 1.29475
4 左 J5 位置 (rad) -1.46910 1.10481 左 J5 目标 (rad) -1.52171 0.96641
5 左 J6 位置 (rad) -2.39948 1.01569 左 J6 目标 (rad) -2.41142 1.01856
6 左夹爪位置 (mm) 0.01054 89.77434 左夹爪目标 (mm) 0.00000 90.00000
7 右 J1 位置 (rad) -1.06315 1.07065 右 J1 目标 (rad) -1.06612 1.07072
8 右 J2 位置 (rad) -0.01895 2.39944 右 J2 目标 (rad) -0.00460 2.46664
9 右 J3 位置 (rad) -1.96907 0.03019 右 J3 目标 (rad) -1.96963 0.01227
10 右 J4 位置 (rad) -0.75332 1.39413 右 J4 目标 (rad) -0.88511 1.26707
11 右 J5 位置 (rad) -1.42548 0.80130 右 J5 目标 (rad) -1.44654 0.87744
12 右 J6 位置 (rad) -3.06723 0.64789 右 J6 目标 (rad) -3.08177 1.20724
13 右夹爪位置 (mm) 0.14510 89.87153 右夹爪目标 (mm) 0.00000 90.00000

右臂 action 的 J4/J5(dim 10/11)在转换时执行了符号修正,以统一 raw action 与机器人 state 的方向约定。原始数据未被修改。

时间对齐

raw 数据中的 state、action、session 事件共享进程内单调时钟:

  • robot_state.csv 提供 state/feedback 时间戳;
  • action.csv 提供 action 时间戳;
  • teleop_events.csv 提供各相机的 VideoRecordStart/Stop
  • video/manifest.json 提供 camera slot、设备和视频段信息。

导出时建立 30 Hz 固定时间网格,state/action 线性插值到对应时刻,视频按相机启动锚点选择最近的帧。每个 episode 的原始时间范围、视频锚点和网格起止时间保存在 meta/ragtel/alignment.v1.json

当前 camera_offsets_ms=0action_delay_ms=0 是导出默认值,并非测量得到的物理延迟。raw 没有逐视频帧的 host 单调时间戳,因此本版本不能被描述为硬件级帧同步;用于精确闭环 replay 前应进一步标定视频延迟和 action delay。

相机标定与 URDF

meta/ragtel_calibration/ 包含:

  • 三路相机内参;
  • 左右腕相机手眼外参;
  • 顶部相机相对左右机械臂 base 的外参;
  • piper_x_w_gripper.urdf 的数据集快照;
  • 标定来源、角色映射和 SHA256 信息。

当前标定按相机角色从参考标定快照重映射到 left_wrist/top/right_wrist。raw manifest 只保存 /dev/video*,没有稳定相机序列号,因此 physical_camera_serial_verified=false。在度量级三维重投影、Overlay 验证或更换相机后,应先核对物理相机序列号,必要时重新标定。

Episode 清单

Episode Raw session 帧数 时长 (s)
0 8_19_2 4,987 166.23
1 8_19_3 4,052 135.07
2 8_19_4 3,494 116.47
3 8_19_6 4,646 154.87
4 8_19_7 3,945 131.50
5 8_19_8 3,685 122.83
6 8_19_9 4,395 146.50
7 8_19_10 3,859 128.63
8 8_19_11 6,062 202.07
9 8_19_12 4,711 157.03
10 8_19_13 3,596 119.87
11 8_19_14 3,753 125.10

数据质量与使用注意事项

  1. 当前发布只包含时长不少于 30 秒且被选入首批转换的 12 条 session。
  2. 三路视频都是真实画面,不包含 dummy/黑色视频;左右臂 14 个数值维度也都是真实数据。
  3. stats.json 提供全量 min/max/mean/std 及分位数,可用于归一化;部署时仍需使用机器人自身限位和安全控制。
  4. 在训练前建议逐 episode 检查画面、state/action tracking error、任务成功状态以及视频时间偏移。
  5. 这是数据集快照。raw 来源和转换 provenance 见 meta/ragtel/batch_conversion.v1.json,不要通过修改 Parquet/MP4 来修正 raw;修订应生成新的版本。
Downloads last month
26