This dataset was created using LeRobot.
Dataset Description
leapshared/sciedu_bench_v2_sim_props8_electroscope_rod
SciEdu-Bench(과학 실험 시뮬 벤치) 학습 시연이다. 실험 하나(electroscope_rod)의 순위 과제마다 시뮬레이터(MuJoCo)에서 지은 시연을 담았다. 평가는 꾸러미(config/bench/sciedu_bench_package.yaml)의 고정 시작 50개로 따로 한다. 이 데이터에는 평가 시작이 없다.
검전기 막대 장면(props-8: 흰 털 상자는 책상에 붙임, 흰 거치대, 검은 대전 막대). 과제 여섯: 막대 집기, 집기 겉모습 갈래(띠 색 빨강·주황·파랑, 장면 기본 자리), 댔다 뗐다, 털에 문지르기, 거치대에 내려놓기, 실험 전체(대본 단계마다 문장이 바뀐다). 막대는 검전기에 닿지 않는다(가까이 갔다 떨어진다, 유도). 닿거나 검전기를 밀면 판정에서 실패라 이 데이터에 없다. 동작마다 가슴 앞 대기 자세로 끝난다.
출처
- 장면 판
props-8, 규격 판2, 꾸러미 판0.7.0-draft(셋 중 하나라도 다르면 다른 데이터다). - 실험 현상 층을 켠 그림이다(검전기 박). 카메라는 물리 시계(1배, 규격
phenomena.observation_time_scale)로 그려 실물 녹화가 본 만큼만 보인다. 사람에게 보여 주는 가속 시계는 쓰지 않았다. 편 끝의 현상 값은sim_episodes.jsonl의phenomena에 있다. - 시작 자세의 출처 녹화:
leapshared/Electroscope_rod_steps_20261106@cfba0089f00b4de485b6271a678b6524ffe16ba0. - 생성 코드: 저장소 커밋
d2fab3e572277d6e767f5a9eb20424eb28f62123, 도구scripts/sim/bench_train_data.py, 짓는 법config/bench/sciedu_bench_train_data.yaml.
생성기
녹화를 재생해 만든 편은 없다(녹화 재생은 학습 데이터 출처가 아니다, 2026-11-06). 편마다 만든 법:
- 장면에서 짠 IK 코드: 300편
판정(평가와 같은 술어)을 통과한 편만 담았다. 행동은 생성기가 보낸 그대로이고, 같은 시작에서 같은 행동을 카메라를 켜고 다시 돌려 그림을 넣었다(판정과 길이가 같은 것만).
과제와 편수
| 과제 | 묶음 | 문장(task) |
편/목표 | 만든 법 | 프레임 |
|---|---|---|---|---|---|
electroscope_rod_pick |
spatial | pick up the black charging rod | 50/50 | 장면에서 짠 IK 코드 50 | 12126 |
electroscope_rod_pick_look |
object | pick up the black charging rod | 50/50 | 장면에서 짠 IK 코드 50 | 12154 |
electroscope_rod_near |
goal | move the rod toward and away from the electroscope | 50/50 | 장면에서 짠 IK 코드 50 | 11905 |
electroscope_rod_rub |
goal | rub the rod on the fur | 50/50 | 장면에서 짠 IK 코드 50 | 16918 |
electroscope_rod_put |
goal | put the rod down on the white holder | 50/50 | 장면에서 짠 IK 코드 50 | 10837 |
electroscope_rod_script |
long | pick up the black charging rod | 50/50 | 장면에서 짠 IK 코드 50 | 135197 |
합 300편, 199137프레임, 30 fps. 시작을 뽑은 녹화 에피소드 수: electroscope_rod_pick 25, electroscope_rod_pick_look 25, electroscope_rod_near 30, electroscope_rod_rub 30, electroscope_rod_put 30, electroscope_rod_script 25.
평가 시작과의 분리
- 규칙: 평가 시작은 그 단계 녹화 에피소드(세션 2부터)를 정렬한 짝수 번째와 소품 씨앗 = 뽑기, 학습 시작은 홀수 번째와 씨앗 10000 + 뽑기
- 잰 것(
scripts/sim/bench_train_data.py check): 같은 과제의 평가 시작과 녹화 에피소드가 겹치는 편 0, 평가 시작과 같은 장면(소품 0.5 cm·팔 5.0도 안) 0, 같은 실험 다른 과제의 평가 시작과 녹화 에피소드를 나누는 편 0(같은 녹화의 다른 단계). - 학습 시작에서 가장 가까운 평가 시작까지 물체 x·y 거리:
electroscope_rod_pick가장 가까운 것 0.84 cm·가운데 2.36 cm;electroscope_rod_pick_look가장 가까운 것 0.0 cm·가운데 0.0 cm;electroscope_rod_near가장 가까운 것 0.84 cm·가운데 4.25 cm;electroscope_rod_rub가장 가까운 것 0.84 cm·가운데 4.25 cm;electroscope_rod_put가장 가까운 것 0.84 cm·가운데 4.27 cm;electroscope_rod_script가장 가까운 것 0.84 cm·가운데 2.36 cm.
꼴
실물 녹화(LeRobot v3.0)와 같다: 30 fps, action·observation.state 16값(팔로워 관절 도, 왼팔 7·왼 그리퍼·오른팔 7·오른 그리퍼, 절대 관절값), 카메라 셋(observation.images.follower_d455f 가슴, left_wrist, right_wrist, 480×640, AV1). robot_type은 openarm_sciedu_sim(꾸러미 io.robot_type_sim)이라 실험 데이터셋끼리 LeRobot 합치기로 하나가 된다. task는 평가가 주는 규격 문장 그대로다(머리말 없음). 실물 녹화의 문장은 앞에 "Do the ... experiment. "가 붙는다.
에피소드마다 과제·만든 법·시작(녹화 에피소드·흔들기·소품 자리)·판정은 sim_episodes.jsonl에 있다.
- 끝 멈춤(행동이 마지막 값에 머문 시간) 가운데: 시연 2.53~3.53초, 출처 녹화 2.5초. 판정을 통과한 틱 뒤에 마지막 명령으로 2.50초(출처 녹화 끝 멈춤의 가운데값) 더 서 있는 틱을 담았다. 그동안 물건이 1 cm 넘게 가거나 5도 넘게 돌거나 쥔 것을 놓친 편은 버렸다(
sim_episodes.jsonl의end_hold).
라이선스
apache-2.0(허브 메타데이터 apache-2.0, 선생님 2026-11-11, 실물 녹화 leapshared/SciEdu_mineral_steps_20261109와 같게). 이 시뮬 시연 데이터셋의 라이선스다. 코드·장면(소품 CAD·사진 결)의 라이선스는 꾸러미의 license 칸에서 따로 정한다.
판(리비전) 고정
학습·평가 기록에는 이 데이터를 저장소@커밋으로 적는다(꾸러미 training_data.sim이 커밋을 고정한다). 부를 때도 커밋을 준다:
lerobot-train --dataset.repo_id=leapshared/sciedu_bench_v2_sim_props8_electroscope_rod --dataset.revision=<커밋> ...
from lerobot.datasets.lerobot_dataset import LeRobotDataset
ds = LeRobotDataset("leapshared/sciedu_bench_v2_sim_props8_electroscope_rod", revision="<커밋>")
장면·판정·시작이 바뀌면 같은 저장소에 덮어쓰지 않고 장면 판이 이름에 든 새 저장소로 올린다. 다시 짓기: scripts/sim/bench_train_data.py all --dataset electroscope_rod.
Training demonstrations of SciEdu-Bench (MuJoCo), experiment electroscope_rod, scene props-8, spec 2, package 0.7.0-draft. 300 episodes made in simulation by scene_code (300); no recording replay. Only judged successes are kept. The training starts are disjoint from the package's fixed evaluation starts (same-task recorded-episode overlap 0, starts within 0.5 cm and 5.0 degrees of one 0). Schema: the real recordings' (30 fps, 16 absolute joint degrees, three 480x640 AV1 cameras); task is the benchmark sentence without the recordings' head. License: apache-2.0 (metadata apache-2.0). Drawn with the phenomenon layer on (electroscope_rod) on the physical clock (1x, spec phenomena.observation_time_scale): the cameras see what the real recordings saw, not the sped-up show clock; each episode's end values are in sim_episodes.jsonl phenomena. Pin a commit (--dataset.revision). Electroscope rod scene (props-8): pick the rod, the same with three band colours, toward and away from the electroscope, rub on the fur, put it back on the holder, and the whole script whose sentence changes step by step. The rod never touches the electroscope (a touch fails the judge). Each motion ends at the standby pose in front of the chest.
- Homepage: [More Information Needed]
- Paper: [More Information Needed]
- License: apache-2.0
Dataset Structure
{
"codebase_version": "v3.0",
"fps": 30,
"features": {
"action": {
"dtype": "float32",
"shape": [
16
],
"names": [
"left_joint_1.pos",
"left_joint_2.pos",
"left_joint_3.pos",
"left_joint_4.pos",
"left_joint_5.pos",
"left_joint_6.pos",
"left_joint_7.pos",
"left_gripper.pos",
"right_joint_1.pos",
"right_joint_2.pos",
"right_joint_3.pos",
"right_joint_4.pos",
"right_joint_5.pos",
"right_joint_6.pos",
"right_joint_7.pos",
"right_gripper.pos"
]
},
"observation.state": {
"dtype": "float32",
"shape": [
16
],
"names": [
"left_joint_1.pos",
"left_joint_2.pos",
"left_joint_3.pos",
"left_joint_4.pos",
"left_joint_5.pos",
"left_joint_6.pos",
"left_joint_7.pos",
"left_gripper.pos",
"right_joint_1.pos",
"right_joint_2.pos",
"right_joint_3.pos",
"right_joint_4.pos",
"right_joint_5.pos",
"right_joint_6.pos",
"right_joint_7.pos",
"right_gripper.pos"
]
},
"observation.images.follower_d455f": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 480,
"video.width": 640,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"video.channels": 3,
"has_audio": false,
"video.g": 2,
"video.crf": 30,
"video.preset": 12,
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": {},
"is_depth_map": false
}
},
"observation.images.left_wrist": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 480,
"video.width": 640,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"video.channels": 3,
"has_audio": false,
"video.g": 2,
"video.crf": 30,
"video.preset": 12,
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": {},
"is_depth_map": false
}
},
"observation.images.right_wrist": {
"dtype": "video",
"shape": [
480,
640,
3
],
"names": [
"height",
"width",
"channels"
],
"info": {
"video.height": 480,
"video.width": 640,
"video.codec": "av1",
"video.pix_fmt": "yuv420p",
"video.fps": 30,
"video.channels": 3,
"has_audio": false,
"video.g": 2,
"video.crf": 30,
"video.preset": 12,
"video.fast_decode": 0,
"video.video_backend": "pyav",
"video.extra_options": {},
"is_depth_map": false
}
},
"timestamp": {
"dtype": "float32",
"shape": [
1
],
"names": null
},
"frame_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"episode_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"index": {
"dtype": "int64",
"shape": [
1
],
"names": null
},
"task_index": {
"dtype": "int64",
"shape": [
1
],
"names": null
}
},
"total_episodes": 300,
"total_frames": 199137,
"total_tasks": 6,
"chunks_size": 1000,
"data_files_size_in_mb": 100,
"video_files_size_in_mb": 200,
"data_path": "data/chunk-{chunk_index:03d}/file-{file_index:03d}.parquet",
"video_path": "videos/{video_key}/chunk-{chunk_index:03d}/file-{file_index:03d}.mp4",
"robot_type": "openarm_sciedu_sim",
"splits": {
"train": "0:300"
}
}
Citation
BibTeX:
[More Information Needed]
- Downloads last month
- 161