Episodes Preview openarm_sciedu_sim Visualizer
62 episodes · 30 fps · 3 cameras · 640×480 av1

This dataset was created using LeRobot.

Dataset Description

쓰기 전에(2026-11-06). 대전판(옛 소품) 장면의 시연이고, 녹화 재생 편 27편이 들어 있다(코드 생성 35편). 2026-11-06부터 시연 출처는 코드 생성과 pi0.5 롤아웃이고 녹화 재생은 점검용이다(선생님 결정). 대전 막대 장면의 데이터셋이 나오면 그것으로 대체한다. / Before use (2026-11-06). Demos of the charging-plate scene (the old prop), 27 of them replays (35 code). From 2026-11-06 the demos come from code and pi0.5 rollouts, and replay is a check only (the teacher's decision). The charging-rod scene's dataset will replace this one when it comes.

과학 실험 시뮬 벤치(SciEdu-Bench 2판, MuJoCo, 장면 props-5)의 시연이다. 실물 녹화(LeRobot 0.6.1, v3.0)와 같은 꼴이다: 30 fps, action·observation.state 16값(팔로워 관절 도, 그리퍼는 리더 눈금; 절대 관절값, 상대 액션 아님), 카메라 셋 (follower_d455f 가슴, left_wrist, right_wrist, 480×640). task는 학습 문장 그대로다(머리말 없음). 시연은 두 길로 만들었다: (가) 코드 생성(녹화 손 길을 옮긴 물체에 실어 IK로 풂), (나) 녹화 재생(시작의 출처 녹화 행동). 과제 판정을 통과한 것만 담았다. 에피소드마다 과제·출처(code/replay)·시작 id·흔들기 번호·출처 녹화 에피소드·판정은 sim_episodes.jsonl에 있다. 평가 시작이 쓰는 녹화 에피소드에서는 시연을 만들지 않았다. 올리기 전에 그림을 보았고, 이상하게 본 에피소드(물체가 튐, 시작에서 물체가 겹침)를 뺐으면 까닭은 sim_dropped.json에 있다. 과제마다 50편을 다 채우지 못한 것이 있다 (평가가 쓰지 않은 녹화만 쓰고, 그중 시작을 지을 수 없거나 판정을 통과하지 못한 것을 채우지 않았다).

Simulated demos of the SciEdu-Bench v2 (MuJoCo, props-5 scene), the real recordings' schema; see sim_episodes.jsonl for each episode's task, source (code / replay), start id and judge result.

props-5 장면판(2026-11-06). 옛 판 leapshared/sciedu_bench_v2_sim_props4_electroscope(props-4, 34편)은 지우지 않고 그대로 둔다. 이 판과 다른 것:

  • 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.
  • 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.
  • 털 상자를 책상에 고정했다(실물은 테이프로 붙인다). props-4에서는 녹화 재생의 판이 상자를 밀어, 판을 든 과제의 시연 시작을 하나도 짓지 못했다.
  • 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.
  • 내려놓기 판정에 아래 한계(원판의 가장 낮은 점 ≥ 상자 윗면 − 3.5 cm)를 더했다.

시작 고침(2026-11-06, 이 판의 두 번째 커밋). 첫 커밋(4e50f8d6, 60편)에서 바꾼 것:

  • 문지르기·내려놓기·가까이 멈춤의 녹화 재생 14편(에피소드 3744, 4651)을 다시 녹화했다. 이 과제들은 규격상 '가까이' 단계의 녹화 프레임에서 시작하는데, 첫 판은 그 과제 단계 녹화 구간의 첫 명령부터 내어 첫 프레임에 팔이 한 관절 최대 33°를 한 프레임에 옮겨 갔다. 고친 재생은 먼저 시작 자세에서 그 첫 명령까지 코드 생성과 같은 이끎(한 틱 평균 1°, 최대 1.5°, 그리퍼는 그대로)으로 간 뒤 녹화를 낸다. 이끎 틱 수는 sim_episodes.jsonl의 lead_ticks다(345틱). 이끎 구간은 녹화가 아니다: 학습에서 빼려면 에피소드 앞 lead_ticks 프레임을 거른다. 이끎 뒤 행동은 첫 판과 같다. 끝 틱은 판정이 정해 −1+2틱 다르다.
  • 검전기로 가져가기 0 → 2편(에피소드 52·53, 출처 녹화 63·64). 검전기 쥠 찾기를 산화환원처럼 '열린 뒤 처음 닫힘'으로 고쳐, 빈손으로 닫힌 채 시작하는 조각 녹화(60~69)의 쥠을 프레임 0이 아니라 실제로 쥐는 프레임에서 잡는다. 녹화 62에서 만든 한 편은 그림에서 뺐다 (sim_dropped.json: 쥔 판이 한 프레임 사이에 손 안에서 돌아 섬).
  • 다른 46편은 첫 판과 행동·읽기·곁 파일이 같다(곁 파일에 lead_ticks 0 또는 코드 생성 이끎 틱만 더함).

과제별 편수: 대전판 집기 22, 집기(겉모습 갈래) 10, 댔다 뗐다 5, 털에 문지르기 5, 내려놓기 5, 댔다 뗐다 하고 멈추기 5, 검전기로 가져가기 2, 접지 기다리기 2, 홈 4, 검전기 실험 전체 2. 합 62. 평가 시작이 쓰지 않은 녹화가 과제마다 20개뿐이고, 그중 10개는 앞 단계만 담은 조각 녹화라 50편을 채운 과제는 없다.

알아 둘 것: 홈 에피소드 59(첫 판 57)는 판이 상자 위가 아니라 상자 앞 책상에 누운 채 시작한다(녹화 재생 중 앞 단계에서 판이 떨어짐). 홈 과제는 판을 보지 않는다.

props-5 scene (2026-11-06). The older leapshared/sciedu_bench_v2_sim_props4_electroscope (props-4, 34 episodes) stays as it is. What changed since it:

  • 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.
  • 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.
  • 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.
  • 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).
  • Putting the plate down also needs its lowest point no more than 3.5 cm below the box top.

Start fixes (2026-11-06, this dataset's second commit). What changed since the first commit (4e50f8d6, 60 episodes):

  • The 14 replayed rub / put / near-and-stop episodes (37 to 44, 46 to 51) were recorded again. These tasks start at the recorded frame of the 'near' step; the first commit played the task step's recorded span from its first command, so one joint moved up to 33° in the first frame. The fixed replay first goes from the start pose to that first command with the code demo's lead-in (1° a tick on average, 1.5° at most, the grippers kept), then plays the recording. lead_ticks in sim_episodes.jsonl gives the lead-in ticks (3 to 45). The lead-in is not the recording: drop each episode's first lead_ticks frames to train on recorded motion only. The actions after the lead-in are the first commit's; the last tick differs by −1 to +2 (the judge ends the episode).
  • Bring to the electroscope 0 → 2 episodes (52, 53; recordings 63, 64). The electroscope grasp is now found as Redox's is (the first close after the gripper opened), so the short recordings 60 to 69, which start with an empty closed gripper, are grasped at the frame the hand really closes on the plate, not at frame 0. One episode made from recording 62 was left out after looking at the frames (sim_dropped.json: the held plate turns upright in the hand between two frames).
  • The other 46 episodes are the first commit's (actions, readings and notes; lead_ticks added: 0, or the code demo's lead-in).

Episodes per task: pick 22, pick (look variant) 10, near 5, rub 5, put 5, near-and-stop 5, bring 2, wait 2, home 4, the whole experiment 2; 62 in all. Only 20 recordings per task are left unused by the evaluation starts, and 10 of them hold the first steps only, so no task reaches 50.

Note: home episode 59 (57 in the first commit) starts with the plate lying on the desk in front of the box (it fell during an earlier step of the replay); the home task does not look at the plate.

  • Homepage: [More Information Needed]
  • Paper: [More Information Needed]
  • License: apache-2.0

Dataset Structure

meta/info.json:

{
    "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": 62,
    "total_frames": 14938,
    "total_tasks": 8,
    "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": "0:62"
    }
}

Citation

BibTeX:

[More Information Needed]
Downloads last month
90