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

This dataset was created using LeRobot.

Dataset Description

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 한 길로 만들었다: (가) 코드 생성(녹화 손 길을 옮긴 물체에 실어 IK로 풂). 과제 판정을 통과한 것만 담았다. 에피소드마다 과제·출처(code)·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), start id and judge result.

시연 출처는 코드 생성뿐이다(2026-11-06, 파란 용액 150 mL 장면). 이 데이터셋에는 녹화 재생으로 만든 시연이 없다. 선생님 결정이다: "녹화 재생은 왜하는거지? 그냥 학습된 모델을 돌려보거나 룰베이스로 ik로 하는걸로 데이터를 새로 만들던가." 녹화 재생은 시뮬이 실물을 따라가는지 보는 평가 기준선으로만 쓴다. 옛 판 leapshared/sciedu_bench_v2_sim_props4_redox(118편, 코드 86·녹화 재생 32)는 지우지 않고 그대로 둔다.

  • 코드 생성: 녹화 손 길을 옮긴 물체에 실어 IK로 푼다.
  • pi0.5 롤아웃 편은 이 판에 없다. 시연 출처 규칙은 코드 생성과 학습한 정책(pi0.5)의 롤아웃이다. 하지만 pi0.5는 새 용액(150 mL) 장면에서 다시 찍지 않아 뺐다(그날 밤 GPU를 다른 학습이 쓰고 있었다). 옛 용액 장면에서 찍은 pi0.5 11편은 장면이 달라 올리지 않았다.
  • 장면: props-5 규격의 산화환원 장면에서 두 가지를 실물에 맞춘 판이다. 파란 용액이 150 mL다(수면 3.69 cm, 선생님 2026-11-06 "150 mL쯤", 앞 판 89.7 mL). 엔진이 팔 관절 명령을 실물 드라이버처럼 자른다(드라이버 모델의 관절 범위 − 0.05 rad). 규격 이름은 그대로 props-5다.
  • 한 시작에서 시연을 두 편 만들지 않는다. 흔들기가 있는 과제(판 들기·파란 용액 위로·왼쪽 비커에 넣기)는 녹화 에피소드마다 흔들기를 다섯 번까지 뽑았다(옛 판 셋).
  • 이끎 구간은 녹화가 아니다. 코드 생성 편은 시작 자세에서 과제 구간의 첫 자세까지 부드럽게 가는 이끎으로 시작할 수 있다(한 틱 평균 1°, 최대 1.5°, 그리퍼 그대로). 그 틱 수가 sim_episodes.jsonl의 lead_ticks이고, 이 판에서는 담그기 2편(65·65틱)과 왼쪽 비커 위로 3편(48·21·17틱)에만 있고 나머지 143편은 0이다. 시연한 움직임만 학습하려면 에피소드 앞 lead_ticks 프레임을 거른다.
  • 그림을 보고 두 편을 뺐다(sim_dropped.json). 판 들기 1편은 왼손 손가락 끝이 왼쪽 비커에 누운 판 끝을 쳐 판이 튀어 선다(판 최고 0.70 m/s). 홈 1편은 시작에서 왼쪽 비커가 파란 비커 자리에 겹친다(두 비커 중심 1.6 cm). 둘 다 옛 판에서 뺀 것과 같은 시작이다.

과제별 편수: 판 들기 49, 파란 용액 위로 42, 왼쪽 비커에 넣기 20, 파란 용액 위로(겉모습 갈래) 9, 판 가리키기 2, 담그기 2, 왼쪽 비커 위로 3, 꺼내기 5, 다시 가리키기 2, 홈 13, 산화환원 실험 전체 1. 합 148편(모두 코드 생성). 옛 판과 견주면 과제별로 49 → 49, 29 → 42, 8 → 20, 9 → 9, 2 → 2, 1 → 2, 3 → 3, 3 → 5, 0 → 2, 13 → 13, 1 → 1이다(118 → 148). 평가 시작이 쓰지 않은 녹화만 써서 50편에 닿은 과제는 판 들기(그림을 보고 하나를 빼서 49) 하나뿐이다.

Demo source: code only (2026-11-06, the 150 mL blue solution scene). No demo here replays a recording. The teacher decided so: why replay the recordings, when we can run the trained model or make new data with rule-based IK. The replay stays as the evaluation's check that the simulation follows the robot. The older leapshared/sciedu_bench_v2_sim_props4_redox (118 episodes: 86 code, 32 replay) stays as it is.

  • Code: the recorded hand path carried onto the moved objects and solved by IK.
  • No pi0.5 rollouts in this version. The source rule is code plus rollouts of the trained policy (pi0.5), but pi0.5 was not run again in the new 150 mL scene (the GPU was busy with another training that night). The 11 pi0.5 demos made in the older liquid scene were not uploaded, as the scene differs.
  • The scene: the Redox scene of the props-5 spec with two changes to match the robot. The blue solution is 150 mL (surface at 3.69 cm, the teacher on 2026-11-06: "about 150 mL"; 89.7 mL before). The engine clamps the arm joint commands as the robot's driver does (the driver model's joint range minus 0.05 rad). The spec name stays props-5.
  • No start gives two demos. The tasks with shifts (lift, move to the blue solution, put in the left beaker) draw up to five shifts per recording (three in the older dataset).
  • The lead-in is not recorded motion. A code demo may begin with a smooth lead-in from the start pose to the first pose of its task span (1° a tick on average, 1.5° at most, grippers kept). lead_ticks in sim_episodes.jsonl counts it; in this version only the 2 dip demos (65 and 65 ticks) and the 3 move-over-the-left-beaker demos (48, 21 and 17 ticks) have one; the other 143 have 0. Drop each episode's first lead_ticks frames to train on the demonstrated motion only.
  • Two demos were left out after looking at their frames (sim_dropped.json). In one lift demo the left fingertips hit the end of the plate lying in the left beaker and the plate jumps upright (plate peak 0.70 m/s). In one home demo the left beaker starts on the blue beaker's spot (centres 1.6 cm apart). Both are the same starts left out of the older dataset.

Episodes per task: lift 49, move to the blue solution 42, put in the left beaker 20, move to the blue solution (look variant) 9, point 2, dip 2, move over the left beaker 3, raise 5, point again 2, home 13, the whole experiment 1; 148 in all (all code). Against the older dataset, task by task: 49 → 49, 29 → 42, 8 → 20, 9 → 9, 2 → 2, 1 → 2, 3 → 3, 3 → 5, 0 → 2, 13 → 13, 1 → 1 (118 → 148). Only recordings the evaluation starts leave unused are used, so only lift (49 after the one left out) reaches 50.

  • 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": 148,
    "total_frames": 26075,
    "total_tasks": 10,
    "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": "0:148"
    }
}

Citation

BibTeX:

[More Information Needed]
Downloads last month
45