leapshared commited on
Commit
4e50f8d
·
verified ·
1 Parent(s): 4ac10a9

Upload README.md with huggingface_hub

Browse files
Files changed (1) hide show
  1. README.md +278 -0
README.md ADDED
@@ -0,0 +1,278 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ ---
2
+ license: apache-2.0
3
+ task_categories:
4
+ - robotics
5
+ tags:
6
+ - LeRobot
7
+ - sciedu-bench
8
+ - simulation
9
+ - mujoco
10
+ - openarm
11
+ configs:
12
+ - config_name: default
13
+ data_files: data/*/*.parquet
14
+ ---
15
+
16
+ This dataset was created using [LeRobot](https://github.com/huggingface/lerobot).
17
+
18
+
19
+ <a class="flex" href="https://huggingface.co/spaces/lerobot/visualize_dataset?path=leapshared/sciedu_bench_v2_sim_props5_electroscope">
20
+ <img class="block dark:hidden" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl.svg"/>
21
+ <img class="hidden dark:block" src="https://huggingface.co/datasets/huggingface/badges/resolve/main/visualize-this-dataset-xl-dark.svg"/>
22
+ </a>
23
+
24
+
25
+ ## Dataset Description
26
+
27
+ 과학 실험 시뮬 벤치(SciEdu-Bench 2판, MuJoCo, 장면 props-5)의 시연이다. 실물 녹화(LeRobot 0.6.1, v3.0)와 같은 꼴이다:
28
+ 30 fps, `action`·`observation.state` 16값(팔로워 관절 도, 그리퍼는 리더 눈금; 절대 관절값, 상대 액션 아님), 카메라 셋
29
+ (`follower_d455f` 가슴, `left_wrist`, `right_wrist`, 480×640). `task`는 학습 문장 그대로다(머리말 없음).
30
+ 시연은 두 길로 만들었다: (가) 코드 생성(녹화 손 길을 옮긴 물체에 실어 IK로 풂), (나) 녹화 재생(시작의 출처 녹화 행동).
31
+ 과제 판정을 통과한 것만 담았다. 에피소드마다 과제·출처(code/replay)·시작 id·흔들기 번호·출처 녹화 에피소드·판정은
32
+ `sim_episodes.jsonl`에 있다. 평가 시작이 쓰는 녹화 에피소드에서는 시연을 만들지 않았다. 올리기 전에 그림을 보았고, 이상하게
33
+ 본 에피소드(물체가 튐, 시작에서 물체가 겹침)를 뺐으면 까닭은 `sim_dropped.json`에 있다. 과제마다 50편을 다 채우지 못한 것이 있다
34
+ (평가가 쓰지 않은 녹화만 쓰고, 그중 시작을 지을 수 없거나 판정을 통과하지 못한 것을 채우지 않았다).
35
+
36
+ Simulated demos of the SciEdu-Bench v2 (MuJoCo, props-5 scene), the real recordings' schema; see `sim_episodes.jsonl`
37
+ for each episode's task, source (code / replay), start id and judge result.
38
+
39
+ **props-5 장면판(2026-11-06).** 옛 판 [`leapshared/sciedu_bench_v2_sim_props4_electroscope`](https://huggingface.co/datasets/leapshared/sciedu_bench_v2_sim_props4_electroscope)(props-4, 34편)은 지우지 않고 그대로 둔다. 이 판과 다른 것:
40
+
41
+ - 털가죽을 털 상자 오른끝(가슴 카메라에서 본 오른쪽)에 맞춰 옮겼다(상자 가운데에서 −2.0 → −3.9 cm). 집기 시작에서 대전판 원판이 털에 걸치지 않는다.
42
+ - 검전기를 1.0 cm 낮췄다(유리 목만 줄임, 뚜껑 윗면이 책상 위 16.04 → 15.04 cm). 녹화 손이 쥔 판이 뚜껑을 파고들지 않게 했다.
43
+ - 털 상자를 책상에 고정했다(실물은 테이프로 붙인다). props-4에서는 녹화 재생의 판이 상자를 밀어, 판을 든 과제의 시연 시작을 하나도 짓지 못했다.
44
+ - 털과 상자 윗면을 무른 접촉으로 두었다(잠정 값). 녹화 손을 따라가면 판 원판이 상자 윗면 밑으로 1~3 cm 눌려, 문지르기 끝 같은 프레임에서 원판이 상자에 묻혀 보인다.
45
+ - 내려놓기 판정에 아래 한계(원판의 가장 낮은 점 ≥ 상자 윗면 − 3.5 cm)를 더했다.
46
+
47
+ 과제별 편수(props-4 → props-5): 대전판 집기 24 → 22, 집기(겉모습 갈래) 9 → 10, 댔다 뗐다 0 → 5, 털에 문지르기 0 → 5, 내려놓기 0 → 5,
48
+ 댔다 뗐다 하고 멈추기 0 → 5, 검전기로 가져가기 0 → 0, 접지 기다리기 0 → 2, 홈 0 → 4, 검전기 실험 전체 1 → 2. 합 34 → 60.
49
+ 평가 시작이 쓰지 않은 녹화가 과제마다 20개뿐이고, 그중 10개는 앞 단계만 담은 조각 녹화라 50편을 채운 과제는 없다.
50
+
51
+ 알아 둘 것:
52
+ - 문지르기·내려놓기·가까이 멈춤의 녹화 재생 14편(에피소드 37~44, 46~51)은 첫 프레임에 팔이 크게 옮겨 간다(한 관절이 한 프레임에 1.7~33°).
53
+ 이 과제들은 규격상 '가까이' 단계의 녹화 프레임에서 시작하고, 녹화 재생은 그 과제 단계 녹화 구간의 첫 명령부터 내기 때문이다. 평가의 녹화 재생과 같은 동작이고, 판은 손에 남는다.
54
+ - 홈 에피소드 57은 판이 상자 위가 아니라 상자 앞 책상에 누운 채 시작한다(녹화 재생 중 앞 단계에서 판이 떨어짐). 홈 과제는 판을 보지 않는다.
55
+
56
+ **props-5 scene (2026-11-06).** The older [`leapshared/sciedu_bench_v2_sim_props4_electroscope`](https://huggingface.co/datasets/leapshared/sciedu_bench_v2_sim_props4_electroscope) (props-4, 34 episodes) stays as it is. What changed since it:
57
+
58
+ - The fur sits at the box's right end (as the chest camera sees it; −2.0 → −3.9 cm from the box centre), so the charging plate no longer rests on it at the pick start.
59
+ - The electroscope is 1.0 cm lower (its glass neck only; cap top 16.04 → 15.04 cm above the desk), so the recorded hand's plate no longer cuts into the cap.
60
+ - The fur box is fixed to the desk (taped down in the real setup). In props-4 the replayed plate pushed the box away and no demo start could be made for the tasks that hold the plate.
61
+ - The fur and the box top are soft contacts (provisional values): following the recorded hand presses the plate 1 to 3 cm below the box top, so the plate looks sunk into the box in some frames (e.g. the end of rubbing).
62
+ - Putting the plate down also needs its lowest point no more than 3.5 cm below the box top.
63
+
64
+ Episodes per task (props-4 → props-5): pick 24 → 22, pick (look variant) 9 → 10, near 0 → 5, rub 0 → 5, put 0 → 5, near-and-stop 0 → 5,
65
+ bring 0 → 0, wait 0 → 2, home 0 → 4, the whole experiment 1 → 2; 34 → 60 in all. Only 20 recordings per task are left unused by the
66
+ evaluation starts, and 10 of them hold the first steps only, so no task reaches 50.
67
+
68
+ Notes: the 14 replayed rub / put / near-and-stop episodes (37 to 44, 46 to 51) move the arm sharply on the first frame (one joint 1.7 to 33° in one frame):
69
+ these tasks start at the recorded frame of the 'near' step and the replay plays the task step's recorded span from its first command, as the
70
+ evaluation's replay does; the plate stays in hand. Home episode 57 starts with the plate lying on the desk in front of the box (it fell during
71
+ an earlier step of the replay); the home task does not look at the plate.
72
+
73
+ - **Homepage:** [More Information Needed]
74
+ - **Paper:** [More Information Needed]
75
+ - **License:** apache-2.0
76
+
77
+ ## Dataset Structure
78
+
79
+ [meta/info.json](meta/info.json):
80
+ ```json
81
+ {
82
+ "codebase_version": "v3.0",
83
+ "fps": 30,
84
+ "features": {
85
+ "action": {
86
+ "dtype": "float32",
87
+ "shape": [
88
+ 16
89
+ ],
90
+ "names": [
91
+ "left_joint_1.pos",
92
+ "left_joint_2.pos",
93
+ "left_joint_3.pos",
94
+ "left_joint_4.pos",
95
+ "left_joint_5.pos",
96
+ "left_joint_6.pos",
97
+ "left_joint_7.pos",
98
+ "left_gripper.pos",
99
+ "right_joint_1.pos",
100
+ "right_joint_2.pos",
101
+ "right_joint_3.pos",
102
+ "right_joint_4.pos",
103
+ "right_joint_5.pos",
104
+ "right_joint_6.pos",
105
+ "right_joint_7.pos",
106
+ "right_gripper.pos"
107
+ ]
108
+ },
109
+ "observation.state": {
110
+ "dtype": "float32",
111
+ "shape": [
112
+ 16
113
+ ],
114
+ "names": [
115
+ "left_joint_1.pos",
116
+ "left_joint_2.pos",
117
+ "left_joint_3.pos",
118
+ "left_joint_4.pos",
119
+ "left_joint_5.pos",
120
+ "left_joint_6.pos",
121
+ "left_joint_7.pos",
122
+ "left_gripper.pos",
123
+ "right_joint_1.pos",
124
+ "right_joint_2.pos",
125
+ "right_joint_3.pos",
126
+ "right_joint_4.pos",
127
+ "right_joint_5.pos",
128
+ "right_joint_6.pos",
129
+ "right_joint_7.pos",
130
+ "right_gripper.pos"
131
+ ]
132
+ },
133
+ "observation.images.follower_d455f": {
134
+ "dtype": "video",
135
+ "shape": [
136
+ 480,
137
+ 640,
138
+ 3
139
+ ],
140
+ "names": [
141
+ "height",
142
+ "width",
143
+ "channels"
144
+ ],
145
+ "info": {
146
+ "video.height": 480,
147
+ "video.width": 640,
148
+ "video.codec": "av1",
149
+ "video.pix_fmt": "yuv420p",
150
+ "video.fps": 30,
151
+ "video.channels": 3,
152
+ "has_audio": false,
153
+ "video.g": 2,
154
+ "video.crf": 30,
155
+ "video.preset": 12,
156
+ "video.fast_decode": 0,
157
+ "video.video_backend": "pyav",
158
+ "video.extra_options": {},
159
+ "is_depth_map": false
160
+ }
161
+ },
162
+ "observation.images.left_wrist": {
163
+ "dtype": "video",
164
+ "shape": [
165
+ 480,
166
+ 640,
167
+ 3
168
+ ],
169
+ "names": [
170
+ "height",
171
+ "width",
172
+ "channels"
173
+ ],
174
+ "info": {
175
+ "video.height": 480,
176
+ "video.width": 640,
177
+ "video.codec": "av1",
178
+ "video.pix_fmt": "yuv420p",
179
+ "video.fps": 30,
180
+ "video.channels": 3,
181
+ "has_audio": false,
182
+ "video.g": 2,
183
+ "video.crf": 30,
184
+ "video.preset": 12,
185
+ "video.fast_decode": 0,
186
+ "video.video_backend": "pyav",
187
+ "video.extra_options": {},
188
+ "is_depth_map": false
189
+ }
190
+ },
191
+ "observation.images.right_wrist": {
192
+ "dtype": "video",
193
+ "shape": [
194
+ 480,
195
+ 640,
196
+ 3
197
+ ],
198
+ "names": [
199
+ "height",
200
+ "width",
201
+ "channels"
202
+ ],
203
+ "info": {
204
+ "video.height": 480,
205
+ "video.width": 640,
206
+ "video.codec": "av1",
207
+ "video.pix_fmt": "yuv420p",
208
+ "video.fps": 30,
209
+ "video.channels": 3,
210
+ "has_audio": false,
211
+ "video.g": 2,
212
+ "video.crf": 30,
213
+ "video.preset": 12,
214
+ "video.fast_decode": 0,
215
+ "video.video_backend": "pyav",
216
+ "video.extra_options": {},
217
+ "is_depth_map": false
218
+ }
219
+ },
220
+ "timestamp": {
221
+ "dtype": "float32",
222
+ "shape": [
223
+ 1
224
+ ],
225
+ "names": null
226
+ },
227
+ "frame_index": {
228
+ "dtype": "int64",
229
+ "shape": [
230
+ 1
231
+ ],
232
+ "names": null
233
+ },
234
+ "episode_index": {
235
+ "dtype": "int64",
236
+ "shape": [
237
+ 1
238
+ ],
239
+ "names": null
240
+ },
241
+ "index": {
242
+ "dtype": "int64",
243
+ "shape": [
244
+ 1
245
+ ],
246
+ "names": null
247
+ },
248
+ "task_index": {
249
+ "dtype": "int64",
250
+ "shape": [
251
+ 1
252
+ ],
253
+ "names": null
254
+ }
255
+ },
256
+ "total_episodes": 60,
257
+ "total_frames": 14526,
258
+ "total_tasks": 8,
259
+ "chunks_size": 1000,
260
+ "data_files_size_in_mb": 100,
261
+ "video_files_size_in_mb": 200,
262
+ "data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
263
+ "video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
264
+ "robot_type": "openarm_sciedu_sim",
265
+ "splits": {
266
+ "train": "0:60"
267
+ }
268
+ }
269
+ ```
270
+
271
+
272
+ ## Citation
273
+
274
+ **BibTeX:**
275
+
276
+ ```bibtex
277
+ [More Information Needed]
278
+ ```