robot-lev commited on
Commit
c4b8e56
·
verified ·
1 Parent(s): 22bad90

Add CC-BY dataset card

Browse files
Files changed (1) hide show
  1. README.md +50 -222
README.md CHANGED
@@ -9,241 +9,69 @@ tags:
9
  - franka
10
  - force-torque
11
  - contact-rich
12
- - LeRobot
13
- configs:
14
- - config_name: default
15
- data_files: data/*/*.parquet
16
  ---
17
 
18
- This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
19
 
 
 
20
 
21
- <a class="flex" href="https://huggingface.co/spaces/lerobot/visualize_dataset?path=robot-lev/fmb">
22
- <img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl.svg"/>
23
- <img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl-dark.svg"/>
24
- </a>
25
 
 
26
 
27
- ## Dataset Description
 
 
 
28
 
 
 
 
 
 
 
29
 
 
30
 
31
- - **Homepage:** [More Information Needed]
32
- - **Paper:** [More Information Needed]
33
- - **License:** cc-by-4.0
 
 
 
 
 
 
 
 
34
 
35
- ## Dataset Structure
 
36
 
37
- [meta/info.json](meta/info.json):
38
- ```json
39
- {
40
- "codebase_version": "v3.0",
41
- "robot_type": "franka",
42
- "total_episodes": 1844,
43
- "total_frames": 418495,
44
- "total_tasks": 7,
45
- "chunks_size": 1000,
46
- "data_files_size_in_mb": 100,
47
- "video_files_size_in_mb": 200,
48
- "fps": 10,
49
- "splits": {
50
- "train": "0:1844"
51
- },
52
- "data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
53
- "video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
54
- "features": {
55
- "observation.state": {
56
- "dtype": "float32",
57
- "shape": [
58
- 28
59
- ],
60
- "names": null
61
- },
62
- "observation.state.joint_position": {
63
- "dtype": "float32",
64
- "shape": [
65
- 7
66
- ],
67
- "names": null
68
- },
69
- "observation.state.ee_pose": {
70
- "dtype": "float32",
71
- "shape": [
72
- 7
73
- ],
74
- "names": null
75
- },
76
- "observation.state.gripper": {
77
- "dtype": "float32",
78
- "shape": [
79
- 1
80
- ],
81
- "names": null
82
- },
83
- "observation.force": {
84
- "dtype": "float32",
85
- "shape": [
86
- 3
87
- ],
88
- "names": null
89
- },
90
- "observation.torque": {
91
- "dtype": "float32",
92
- "shape": [
93
- 3
94
- ],
95
- "names": null
96
- },
97
- "observation.jacobian": {
98
- "dtype": "float32",
99
- "shape": [
100
- 42
101
- ],
102
- "names": null
103
- },
104
- "action": {
105
- "dtype": "float32",
106
- "shape": [
107
- 7
108
- ],
109
- "names": null
110
- },
111
- "observation.images.side_1": {
112
- "dtype": "video",
113
- "shape": [
114
- 256,
115
- 256,
116
- 3
117
- ],
118
- "names": [
119
- "height",
120
- "width",
121
- "channels"
122
- ],
123
- "info": {
124
- "video.height": 256,
125
- "video.width": 256,
126
- "video.codec": "av1",
127
- "video.pix_fmt": "yuv420p",
128
- "video.is_depth_map": false,
129
- "video.fps": 10,
130
- "video.channels": 3,
131
- "has_audio": false
132
- }
133
- },
134
- "observation.images.side_2": {
135
- "dtype": "video",
136
- "shape": [
137
- 256,
138
- 256,
139
- 3
140
- ],
141
- "names": [
142
- "height",
143
- "width",
144
- "channels"
145
- ],
146
- "info": {
147
- "video.height": 256,
148
- "video.width": 256,
149
- "video.codec": "av1",
150
- "video.pix_fmt": "yuv420p",
151
- "video.is_depth_map": false,
152
- "video.fps": 10,
153
- "video.channels": 3,
154
- "has_audio": false
155
- }
156
- },
157
- "observation.images.wrist_1": {
158
- "dtype": "video",
159
- "shape": [
160
- 256,
161
- 256,
162
- 3
163
- ],
164
- "names": [
165
- "height",
166
- "width",
167
- "channels"
168
- ],
169
- "info": {
170
- "video.height": 256,
171
- "video.width": 256,
172
- "video.codec": "av1",
173
- "video.pix_fmt": "yuv420p",
174
- "video.is_depth_map": false,
175
- "video.fps": 10,
176
- "video.channels": 3,
177
- "has_audio": false
178
- }
179
- },
180
- "observation.images.wrist_2": {
181
- "dtype": "video",
182
- "shape": [
183
- 256,
184
- 256,
185
- 3
186
- ],
187
- "names": [
188
- "height",
189
- "width",
190
- "channels"
191
- ],
192
- "info": {
193
- "video.height": 256,
194
- "video.width": 256,
195
- "video.codec": "av1",
196
- "video.pix_fmt": "yuv420p",
197
- "video.is_depth_map": false,
198
- "video.fps": 10,
199
- "video.channels": 3,
200
- "has_audio": false
201
- }
202
- },
203
- "timestamp": {
204
- "dtype": "float32",
205
- "shape": [
206
- 1
207
- ],
208
- "names": null
209
- },
210
- "frame_index": {
211
- "dtype": "int64",
212
- "shape": [
213
- 1
214
- ],
215
- "names": null
216
- },
217
- "episode_index": {
218
- "dtype": "int64",
219
- "shape": [
220
- 1
221
- ],
222
- "names": null
223
- },
224
- "index": {
225
- "dtype": "int64",
226
- "shape": [
227
- 1
228
- ],
229
- "names": null
230
- },
231
- "task_index": {
232
- "dtype": "int64",
233
- "shape": [
234
- 1
235
- ],
236
- "names": null
237
- }
238
- }
239
- }
240
- ```
241
 
 
 
 
 
 
 
242
 
243
  ## Citation
244
 
245
- **BibTeX:**
246
-
247
  ```bibtex
248
- [More Information Needed]
249
- ```
 
 
 
 
 
 
 
 
9
  - franka
10
  - force-torque
11
  - contact-rich
12
+ size_categories:
13
+ - 1M<n<10M
 
 
14
  ---
15
 
16
+ # FMB single-object (LeRobot v3)
17
 
18
+ A LeRobot Dataset v3 port of the **FMB** (Functional Manipulation Benchmark)
19
+ **single-object manipulation** demonstrations, recorded with a Franka Panda arm.
20
 
21
+ > **This is a reformatted derivative**, not the original release. The original data and
22
+ > full documentation are published by the authors:
23
+ > **https://huggingface.co/datasets/charlesxu0124/functional-manipulation-benchmark**
24
+ > Paper: [arXiv:2401.08553](https://arxiv.org/abs/2401.08553) · Project: https://functional-manipulation-benchmark.github.io
25
 
26
+ ## What this is
27
 
28
+ FMB ships one `.npy` per demonstration (4 RGB + 4 depth cameras, proprioception, 6-axis
29
+ end-effector force/torque, a commanded cartesian action, and per-step skill primitives).
30
+ This port converts each single-object demonstration into **one LeRobot episode**, keeping
31
+ the RGB streams, proprioception, force/torque, and action on a uniform frame grid.
32
 
33
+ - **Episodes:** 1844
34
+ - **Frames:** 418,495 @ 10 fps
35
+ - **Robot:** Franka Panda
36
+ - **Cameras:** `side_1`, `side_2`, `wrist_1`, `wrist_2` (RGB 256×256)
37
+ - **Per-frame task:** the active skill primitive (e.g. *grasp*, *insert*, *rotate*)
38
+ - **Scope:** single-object subset only (FMB's multi-object subset is not included in this port).
39
 
40
+ ## Features
41
 
42
+ | key | dtype | shape | notes |
43
+ |---|---|---|---|
44
+ | `observation.images.{side_1,side_2,wrist_1,wrist_2}` | video | 256×256×3 | RGB (converted from FMB's BGR) |
45
+ | `observation.state` | float32 | (28,) | joint pos (7) + joint vel (7) + EE pose (7) + EE vel (6) + gripper (1) |
46
+ | `observation.state.joint_position` | float32 | (7,) | |
47
+ | `observation.state.ee_pose` | float32 | (7,) | xyz + quaternion, base frame |
48
+ | `observation.state.gripper` | float32 | (1,) | 0=open, 1=closed |
49
+ | `observation.force` | float32 | (3,) | end-effector force, **EE frame** |
50
+ | `observation.torque` | float32 | (3,) | end-effector torque, **EE frame** |
51
+ | `observation.jacobian` | float32 | (42,) | robot jacobian (6×7), flattened |
52
+ | `action` | float32 | (7,) | commanded cartesian: xyz, rpy, gripper |
53
 
54
+ Per-episode object metadata (shape/size/length/color/angle/distractor + `object_info`) is in
55
+ `meta/fmb_episodes.json`.
56
 
57
+ ## Fidelity notes (please read)
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
58
 
59
+ - **Depth dropped.** FMB's 4 depth maps are **not** included in this port (RGB + F/T + proprio
60
+ + action only). Use the original dataset if you need depth.
61
+ - **BGR → RGB.** FMB stores images in BGR; they are converted to RGB here.
62
+ - **Action is the FMB commanded action as-is** (no next-pose reconstruction).
63
+ - **fps = 10 is nominal.** The source `.npy` carry no timestamps; frames map 1:1, so `fps`
64
+ is metadata, not a resampling rate.
65
 
66
  ## Citation
67
 
 
 
68
  ```bibtex
69
+ @article{luo2024fmb,
70
+ title = {FMB: a Functional Manipulation Benchmark for Generalizable Robotic Learning},
71
+ author = {Luo, Jianlan and Xu, Charles and Liu, Fangchen and Tan, Liam and Lin, Zipeng and Wu, Jeffrey and Abbeel, Pieter and Levine, Sergey},
72
+ journal = {arXiv preprint arXiv:2401.08553},
73
+ year = {2024}
74
+ }
75
+ ```
76
+
77
+ Conversion scripts: https://github.com/lvjonok/fmb-lerobot-port