so101_sim_v4 — SO-101 sim pick-and-place (clean timing)

312 episodes, 408,910 frames, 30 fps, LeRobot v3 (robot_type: so_follower). Recorded 2026-10-08 in Isaac Sim with the fixed 30 Hz control clock: no repeated or stuttering actions (older datasets repeated the commanded action one frame in three).

Subset	Episodes	What
no_intervention	218	clean demos: the robot picks a pen and puts it in the cup (pens in shuffled order), plus LLM-planned tasks
human_corrected	94	the robot starts 2 cm off (left or right), a simulated person corrects it by voice, the robot finishes
train / val	281 / 31	random split

The episode index lists are in meta/splits.json. Pick the HIL share at training time, e.g. --dataset.episodes='[...]' with all no_intervention episodes and part of human_corrected.

• Features: observation.state, action (6 joints, leader units: body joints in degrees, gripper 0–100), observation.images.top, observation.images.wrist (640×480 AV1 video). For SmolVLA: --rename_map '{"observation.images.top": "observation.images.camera1", "observation.images.wrist": "observation.images.camera2"}'.
• Who was in control, per frame: meta/hil_frames.jsonl: hil.control_source 0 = nobody (idle), 1 = robot policy, 2 = voice correction, 3 = teleop; plus hil.intervening, hil.intervention_level, hil.phase (code tables in meta/conversion.json and meta/hil_phases.json). Use it to drop or down-weight the robot's pre-intervention frames.
• Per episode: meta/hil_episodes.jsonl (source episode, outcome, injected offset, interventions, layout).
• Known limits: sim images (perfect lighting); joint units come from the team's leader-to-sim map, which differs from the real follower's own calibration by up to ~9° (shoulder_pan): measure the real arm before running a sim-trained policy on it.

한국어 요약

• 312 에피소드 / 408,910 프레임 / 30 fps, LeRobot v3. 2026-10-08에 제어 타이밍 버그를 고친 코드로 새로 녹화했습니다 (기존 데이터의 끊기는 액션 없음).
• no_intervention 218개 (깨끗한 데모), human_corrected 94개 (2 cm 틀어진 상태에서 음성으로 수정). 에피소드 목록은 meta/splits.json에 있어서 학습할 때 HIL 비율을 직접 정할 수 있습니다.
• 프레임마다 제어 주체(로봇/사람)는 meta/hil_frames.jsonl에 있습니다. 개입 직전 구간을 빼거나 가중치를 낮출 때 사용하세요.
• 주의: 시뮬레이션 영상이고, 관절 단위가 실제 follower 캘리브레이션과 최대 약 9° 차이가 날 수 있습니다 (shoulder_pan). 실제 로봇에 쓰기 전에 확인이 필요합니다.