Dataset Viewer
Auto-converted to Parquet Duplicate
file
string
structural_checks_passed
bool
errors
list
samples
int64
sampled_duration_seconds
float64
observed_sample_hz
float64
largest_sample_gap_seconds
float64
timing_checks_passed
bool
master_policy
string
master_hold_statistics
dict
task
string
status
string
quality_notes
list
/home/songlin/nb666/data/episode-20261016-181519-809496/episode.hdf5
true
[]
387
38.599817
10.000047
0.102962
true
hold
{ "left": { "held_joint_samples": 152, "held_gripper_samples": 152, "max_joint_command_age_ms": 5939.159161 }, "right": { "held_joint_samples": 151, "held_gripper_samples": 151, "max_joint_command_age_ms": 5684.289387 } }
将木块堆叠起来
complete
[ "Master joints are observed CAN control packets, not independently measured master encoders; master EE is FK-derived.", "In explicit hold mode, repeated last-known commands are marked with original RX times and ages. They are not fresh master-state measurements.", "Software pairs host reception times; no cross-...

Cobot Magic HW2: 木块堆叠示教

本数据集用于《机器人具身智能系统开发实践基础》Part 2 数据采集作业。 包含一条任务为“将木块堆叠起来”的双臂遥操作轨迹,共 387 个采样点, 首末采样间隔约 38.6 秒,采样率约 10 Hz。

文件组织

  • raw/episode.hdf5:原始 RGB、深度、机器人数据和时间戳。
  • raw/metadata.json:参数来源、坐标系、相机内参及标定快照。
  • raw/summary.json、raw/validation.json:原始采集统计和结构检查。
  • MANIFEST.json:上传文件的大小及 SHA-256。

四个原始采集文件保持字节不变,其大小和 SHA-256 已与最初从 Songlin 传输至 Mac 的清单逐项核对。元数据中的路径和设备映射属于当时现场环境。

HDF5 数据字段

字段 内容
Task UTF-8 标量字符串:将木块堆叠起来
observations/rgb/{main,left_wrist,right_wrist} uint8 RGB,形状 (387,480,640,3)
observations/depth_raw/{main,left_wrist,right_wrist} uint16 Z16 深度,形状 (387,480,640)
observations/slave/{left,right}/joint_position_rad 从臂六轴实测关节角,弧度
observations/slave/{left,right}/T_base_tool 从臂反馈转换得到的 4x4 末端位姿,平移单位米
observations/master/{left,right}/joint_command_rad 监听到的六轴控制指令,弧度
observations/master/{left,right}/T_base_tool_fk 控制关节角经过 SDK FK 推算的末端位姿
action/{left,right} 六轴控制关节角及夹爪开度,共七维
timestamps/ 原始相机时间戳、CAN 接收时间和主机采样时间
quality/ 相机到达时间跨度、指令年龄及保持标记

RGB 与原始深度未对齐,应分别使用各自内参。深度米值等于原始数值乘以 metadata 中的 depth_scale_m_per_unit(约 0.001);零值表示无有效深度。 内参、深度到彩色的变换以及三个手眼标定结果同时嵌入 HDF5 的 metadata_json 属性。主相机手眼外参相对于右臂基座;腕部外参相对于各自末端。

读取示例

安装 h5py、numpy 和 huggingface_hub 后:

import json
import h5py
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    repo_id="HappySemt0/cobot-magic-hw2-block-stacking",
    filename="raw/episode.hdf5", repo_type="dataset",
)
with h5py.File(path, "r") as f:
    print(f["Task"].asstr()[()])
    metadata = json.loads(f.attrs["metadata_json"])
    rgb = f["observations/rgb/main"][0]
    depth = f["observations/depth_raw/main"][0]
    depth_m = depth.astype("float32") * metadata["cameras"]["main"]["depth_scale_m_per_unit"]
    joints = f["observations/slave/right/joint_position_rad"][:]
    poses = f["observations/slave/right/T_base_tool"][:]

质量检查与限制

原始逐帧审计通过全部 RGB/Depth 读取、数据长度、有限数值、帧号及时间戳检查; 采样最大间隔约 103 ms。主机按接收时间关联数据,未实施跨相机硬件曝光同步。

主臂字段是 CAN 控制目标及 FK 估计,并非独立读取的主臂编码器或实测末端位姿。 采用显式 hold 策略:左臂 152/387、右臂 151/387 个样本保持上次已接收的指令, 最大指令年龄分别约 5.94、5.68 秒。原接收时间、年龄和保持标记全部保留。 SDK FK 模型与固件的一致性、绝对定位精度未独立验证。

左腕 20 帧、右腕 12 帧的非零深度覆盖率低于 20%,主相机为 0 帧。 该比例只衡量深度可用性,不表示测距精度;本版本保留完整原始轨迹,未删帧。 主相机标定快照记录了未解决的 RGB-PnP 与深度尺度差异(比例约 1.164), 使用绝对三维位置时应考虑该限制。

本数据集包含一条示教,不提供预定义训练/测试划分;Dataset Card 不是完整课程实验报告。

Downloads last month
60