Hnyquister commited on
Commit
546a116
·
verified ·
1 Parent(s): cbc93c3

Translate dataset card to English

Browse files
Files changed (1) hide show
  1. README.md +45 -46
README.md CHANGED
@@ -3,7 +3,6 @@ pretty_name: Tower-SimData
3
  license: cc-by-4.0
4
  language:
5
  - en
6
- - zh
7
  tags:
8
  - robotics
9
  - bimanual-manipulation
@@ -15,11 +14,11 @@ tags:
15
 
16
  # Tower-SimData
17
 
18
- Tower-SimData 是面向双臂 Jenga 操作研究的仿真示教数据集。每个 Episode 保存为一个独立 HDF5,包含双臂状态与命令、末端位姿、四路 JPEG 相机和仿真任务状态。
19
 
20
- 数据集主页:<https://huggingface.co/datasets/tower-benchmark/TOWER-SimData>
21
 
22
- ## 数据规模
23
 
24
  | Task | Name | Public folder | Episodes | Frames | Size |
25
  |---|---|---|---:|---:|---:|
@@ -31,9 +30,9 @@ Tower-SimData 是面向双臂 Jenga 操作研究的仿真示教数据集。每
31
  | Task3-3 | Vertical Stacking, 3 layers | `task3-3_vertical_stacking_3` | 55 | 140,900 | 32.55 GiB |
32
  | **Total** | | | **331** | **983,758** | **225.27 GiB** |
33
 
34
- 公开版本不预设训练集和测试集,331 个 Episode 统一发布。使用者应根据自己的实验协议划分数据并记录划分方式。
35
 
36
- ## 目录
37
 
38
  ```text
39
  data/
@@ -49,47 +48,47 @@ CHECKSUMS.sha256
49
  schema/tower_simdata_hdf5.json
50
  ```
51
 
52
- `MANIFEST.jsonl`记录每个Episode的相对路径、帧数、文件大小、SHA256、采集日期和Schema变体;`CHECKSUMS.sha256`用于下载完整性校验。
53
 
54
- ## 任务定义
55
 
56
- `N`表示任务编号,`L`和`X`都表示积木结构层数。目录统一使用`taskN-L_任务名称`,例如`task1-2_tower_transfer_2`表示 Task1 的2层数据。
57
 
58
- | Task | 中文名称 | 英文目录格式 |
59
  |---|---|---|
60
- | Task1 | 双块塔转移 | `task1-X_tower_transfer_X` |
61
- | Task2 | 十字塔转移 | `task2-X_cross_tower_transfer_X` |
62
- | Task3 | 竖直重叠堆叠 | `task3-X_vertical_stacking_X` |
63
 
64
- ## HDF5核心字段
65
 
66
  | Field | Shape | Description |
67
  |---|---|---|
68
- | `observations/qpos` | `[T,16]` | 双臂实测状态:每臂7个关节加1个夹爪 |
69
- | `leader/qpos` | `[T,16]` | 双臂绝对控制命令:每臂7个关节加1个夹爪 |
70
- | `observations/gripper_status` | `[T,8]` | 两个夹爪的扩展状态,每个夹爪4维 |
71
- | `observations/eef` | `[T,14]` | 双臂末端位姿,`xyz + quaternion(xyzw)` |
72
- | `leader/eef` | `[T,14]` | 控制端末端位姿 |
73
- | `observations/lifter/position_mm` | `[T,2]` | 左、后升降平台的位置读数 |
74
- | `observations/lifter/physical_height_mm` | `[T,2]` | 左、后升降平台的物理高度 |
75
- | `observations/lifter/is_moving` | `[T,2]` | 升降平台是否正在运动 |
76
- | `observations/lifter/valid` | `[T,2]` | 升降平台观测是否有效 |
77
- | `lifter/action_mm` | `[T,2]` | 左、后升降平台的绝对动作命令 |
78
- | `lifter/active_target_mm` | `[T,2]` | 当前生效的升降目标 |
79
- | `lifter/command_delta_mm` | `[T,2]` | 相邻升降命令的变化量 |
80
- | `observations/images/*` | `[T]` | 四路变长JPEG图像 |
81
- | `sim/time` | `[T]` | 严格递增的仿真时间 |
82
- | `sim/action_valid` | `[T]` | 动作有效标记,终行是`false` |
83
- | `sim/terminal` | `[T]` | Episode终止标记 |
84
- | `sim/success` | `[T]` | 任务成功状态 |
85
-
86
- 双臂顺序固定为`left, back`。相机为`left_back_camera`、`right_front_camera`、`left_wrist_camera`和`back_wrist_camera`。原始采样率为30 FPS。
87
-
88
- 所有任务统一使用`tower_simdata_hdf5` Schema,37个必选字段及可选字段定义见`schema/tower_simdata_hdf5.json`。`keyframes`为通用可选字段,Task2/3 还可包含`sim/finger_pad_centers_m`接触点字段。
89
-
90
- ### 18维Action定义
91
-
92
- 公开HDF5将18维动作拆成两个同步字段保存,而不是额外复制一份`action[T,18]`:
93
 
94
  ```text
95
  action[18] = concat(
@@ -105,11 +104,11 @@ leader/qpos[16] = [
105
  lifter/action_mm[2] = [left_lifter, back_lifter]
106
  ```
107
 
108
- J1–J7单位为弧度,夹爪为连续绝对开合量,升降平台单位为毫米。使用者在构造18维向量时必须按上述顺序拼接,不能对三类物理量直接使用同一量纲解释。
109
 
110
- ## 分任务相机预览
111
 
112
- 每张图使用该任务一个代表Episode,第一行为开始帧、第二行为结束帧;每行依次展示四路相机。
113
 
114
  ### Task1-2 · `task1-2_tower_transfer_2`
115
 
@@ -135,7 +134,7 @@ J1–J7单位为弧度,夹爪为连续绝对开合量,升降平台单位为
135
 
136
  ![Task3-3 start and end](previews/task3-3_vertical_stacking_3_start_end.jpg)
137
 
138
- ## 使用
139
 
140
  ```bash
141
  pip install h5py
@@ -151,8 +150,8 @@ PY
151
  sha256sum --check CHECKSUMS.sha256
152
  ```
153
 
154
- 数据保留原始30 FPS时序,未应用面向特定模型的动作变换或归一化。
155
 
156
- ## License and citation
157
 
158
- The dataset is released under the [Creative Commons Attribution 4.0 International License](LICENSE). If you use it, cite the metadata in [`CITATION.cff`](CITATION.cff).
 
3
  license: cc-by-4.0
4
  language:
5
  - en
 
6
  tags:
7
  - robotics
8
  - bimanual-manipulation
 
14
 
15
  # Tower-SimData
16
 
17
+ Tower-SimData is a simulated demonstration dataset for research on bimanual Jenga manipulation. Each episode is stored as an independent HDF5 file containing bimanual robot states and commands, end-effector poses, four JPEG camera streams, and simulation task states.
18
 
19
+ Dataset repository: <https://huggingface.co/datasets/tower-benchmark/TOWER-SimData>
20
 
21
+ ## Dataset Size
22
 
23
  | Task | Name | Public folder | Episodes | Frames | Size |
24
  |---|---|---|---:|---:|---:|
 
30
  | Task3-3 | Vertical Stacking, 3 layers | `task3-3_vertical_stacking_3` | 55 | 140,900 | 32.55 GiB |
31
  | **Total** | | | **331** | **983,758** | **225.27 GiB** |
32
 
33
+ The public release does not define an official training or test split. All 331 episodes are released together. Users should create and document splits that match their experimental protocol.
34
 
35
+ ## Repository Layout
36
 
37
  ```text
38
  data/
 
48
  schema/tower_simdata_hdf5.json
49
  ```
50
 
51
+ `MANIFEST.jsonl` records the relative path, frame count, file size, SHA256 digest, collection date, and schema variant of every episode. `CHECKSUMS.sha256` can be used to verify download integrity.
52
 
53
+ ## Task Definitions and Naming
54
 
55
+ `N` denotes the task index. `L` and `X` denote the number of layers in the block structure. Public folders follow `taskN-L_task_name`; for example, `task1-2_tower_transfer_2` contains the two-layer variant of Task1.
56
 
57
+ | Task | Task name | Folder naming pattern |
58
  |---|---|---|
59
+ | Task1 | Tower Transfer | `task1-X_tower_transfer_X` |
60
+ | Task2 | Cross Tower Transfer | `task2-X_cross_tower_transfer_X` |
61
+ | Task3 | Vertical Stacking | `task3-X_vertical_stacking_X` |
62
 
63
+ ## Core HDF5 Fields
64
 
65
  | Field | Shape | Description |
66
  |---|---|---|
67
+ | `observations/qpos` | `[T,16]` | Recorded bimanual state: seven arm joints and one gripper value per arm |
68
+ | `leader/qpos` | `[T,16]` | Absolute bimanual control commands: seven arm joints and one gripper value per arm |
69
+ | `observations/gripper_status` | `[T,8]` | Extended gripper state, four values per gripper |
70
+ | `observations/eef` | `[T,14]` | Bimanual end-effector poses as `xyz + quaternion(xyzw)` |
71
+ | `leader/eef` | `[T,14]` | Command-side end-effector poses |
72
+ | `observations/lifter/position_mm` | `[T,2]` | Left and back lifter position readings |
73
+ | `observations/lifter/physical_height_mm` | `[T,2]` | Physical heights of the left and back lifters |
74
+ | `observations/lifter/is_moving` | `[T,2]` | Whether each lifter is moving |
75
+ | `observations/lifter/valid` | `[T,2]` | Whether each lifter observation is valid |
76
+ | `lifter/action_mm` | `[T,2]` | Absolute commands for the left and back lifters |
77
+ | `lifter/active_target_mm` | `[T,2]` | Active target of each lifter |
78
+ | `lifter/command_delta_mm` | `[T,2]` | Difference between consecutive lifter commands |
79
+ | `observations/images/*` | `[T]` | Four variable-length JPEG camera streams |
80
+ | `sim/time` | `[T]` | Strictly increasing simulation time |
81
+ | `sim/action_valid` | `[T]` | Action-valid flag; the final row is `false` |
82
+ | `sim/terminal` | `[T]` | Episode termination flag |
83
+ | `sim/success` | `[T]` | Task success state |
84
+
85
+ The fixed arm order is `left, back`. The four cameras are `left_back_camera`, `right_front_camera`, `left_wrist_camera`, and `back_wrist_camera`. The source sampling rate is 30 FPS.
86
+
87
+ All tasks use the unified `tower_simdata_hdf5` schema. The 37 required fields and optional fields are defined in `schema/tower_simdata_hdf5.json`. `keyframes` is an optional general-purpose field. Task2 and Task3 episodes may additionally contain the `sim/finger_pad_centers_m` contact-point field.
88
+
89
+ ### 18-Dimensional Action Convention
90
+
91
+ The public HDF5 files store the 18-dimensional action in two synchronized fields instead of duplicating it as an additional `action[T,18]` dataset:
92
 
93
  ```text
94
  action[18] = concat(
 
104
  lifter/action_mm[2] = [left_lifter, back_lifter]
105
  ```
106
 
107
+ J1–J7 are measured in radians, gripper values are continuous absolute opening commands, and lifter commands are measured in millimeters. These quantities must be concatenated in the order above and must not be interpreted as sharing one physical unit.
108
 
109
+ ## Camera Previews
110
 
111
+ Each preview uses one representative episode. The first row shows the initial frame and the second row shows the final frame. Within each row, the four camera views are shown in a fixed order.
112
 
113
  ### Task1-2 · `task1-2_tower_transfer_2`
114
 
 
134
 
135
  ![Task3-3 start and end](previews/task3-3_vertical_stacking_3_start_end.jpg)
136
 
137
+ ## Reading and Verification
138
 
139
  ```bash
140
  pip install h5py
 
150
  sha256sum --check CHECKSUMS.sha256
151
  ```
152
 
153
+ The dataset preserves the original 30 FPS timeline. No model-specific action transformation or normalization has been applied.
154
 
155
+ ## License and Citation
156
 
157
+ Tower-SimData is released under the [Creative Commons Attribution 4.0 International License](LICENSE). If you use this dataset, please cite the metadata in [`CITATION.cff`](CITATION.cff).