Index
2026-09-04 — Dataset

molmo-augment 데이터셋 갤러리: open / close / place

HF Keh0t0/molmo-augment · 학습 카메라(randomized_zed2_analogue_1) + wrist 나란히, 15 fps · open/close 학습 투입 전 점검용

TL;DR

7,516
open 에피소드
7,558
close 에피소드
14,567
place train 클립
60.7 GB
open+close 용량

1 길이 분포와 형식

episode length histograms
Figure 1. 에피소드 길이(66 ms/step). open 중앙값 11.4 s, close 8.4 s, place 클립 6.5 s(표본 845편).
h5 traj_i: actions/{commanded_action, ee_pose, ee_twist, joint_pos, joint_pos_rel} (uint8 zero-pad JSON, 원본과 동일) · obs/agent/qpos (arm 7 + gripper 2) obs_scene: task_type open|close, task_description, referral_expressions.pickup_obj_name [(표현, 확률)…], policy_dt_ms 66 카메라 6종: droid_shoulder_light_randomization · randomized_gopro_analogue_1 · randomized_zed2_analogue_1(학습 exo) · randomized_zed2_analogue_2 · wrist_camera_zed_mini(학습 wrist) · wrist_camera_zed_mini_depth(미사용) 우리 로더용 준비: valid_trajectory_index.json(index.json 기반 생성) · aggregated_stats.json(공식 calculate_stats.py) · prompt_templates에 open/close 추가

2 Open — FrankaOpenSlideOmniCam

서랍(chest of drawers / dresser) 열기. 집당 최대 10편 시도, 성공 편만. 지시문 "Open the {물체}" + referral 표현(다크우드 드레서 등) 확률 목록.

규모: 7,516편 · 1,078집 · 16 shard · 32.7 GB · 길이 중앙값 11.4 s

house_0 — "Open the chest of drawers" 201 steps ≈ 13.3 s

first framelast frame

첫 프레임 · 마지막 프레임 (마지막 프레임 = v2 goal). referral 예: dark wood dresser with three drawers, dark wood dresser, dark dresser, dark wood dresser with drawers · 마지막 step joint 0.37 (target 0.37)

house_11 — "Open the chest of drawers" 173 steps ≈ 11.4 s

first framelast frame

첫 프레임 · 마지막 프레임 (마지막 프레임 = v2 goal). referral 예: chest, chestofdrawers, light dresser, chest of drawers · 마지막 step joint 0.00 (target 0.24)

house_52 — "Open the light wood dresser" 140 steps ≈ 9.2 s

first framelast frame

첫 프레임 · 마지막 프레임 (마지막 프레임 = v2 goal). referral 예: chest of drawers, dresser, light wood dresser, chestofdrawers · 마지막 step joint 0.00 (target 0.21)

house_74 — "Open the modern white dresser with four drawers" 145 steps ≈ 9.6 s

first framelast frame

첫 프레임 · 마지막 프레임 (마지막 프레임 = v2 goal). referral 예: modern white dresser with four drawers, white dresser with drawers, chest of drawers, white dresser · 마지막 step joint 0.21 (target 0.21)

house_81 — "Open the dark wood dresser with three drawers" 196 steps ≈ 12.9 s

first framelast frame

첫 프레임 · 마지막 프레임 (마지막 프레임 = v2 goal). referral 예: dark wood dresser with three drawers, dark wood dresser, dark dresser, dark wood dresser with drawers · 마지막 step joint 0.38 (target 0.38)

3 Close — FrankaCloseSlideOmniCam

열린 서랍(일부 냉장고 문) 닫기. 로봇 배치 반경 0.55–1.1 m. 지시문 "Close the {물체}".

규모: 7,558편 · 1,316집 · 14 shard · 28.0 GB · 길이 중앙값 8.4 s

house_0 — "Close the chest of drawers" 190 steps ≈ 12.5 s

first framelast frame

첫 프레임 · 마지막 프레임 (마지막 프레임 = v2 goal). referral 예: dark wood dresser with three drawers, dark wood dresser, dark dresser, dark wood dresser with drawers

house_10 — "Close the dark dresser with geometric pattern" 124 steps ≈ 8.2 s

first framelast frame

첫 프레임 · 마지막 프레임 (마지막 프레임 = v2 goal). referral 예: dark dresser with geometric pattern, modern dark dresser, dark dresser, furniture

house_11 — "Close the polished wooden dresser with doors" 92 steps ≈ 6.1 s

first framelast frame

첫 프레임 · 마지막 프레임 (마지막 프레임 = v2 goal). referral 예: polished wooden dresser with doors, brown polished dresser with textured doors, furniture, article of furniture

house_39 — "Close the dark gray fridge" 192 steps ≈ 12.7 s

first framelast frame

첫 프레임 · 마지막 프레임 (마지막 프레임 = v2 goal). referral 예: large dark gray metallic fridge, dark gray fridge, dark fridge, large dark gray fridge

house_65 — "Close the furniture" 155 steps ≈ 10.2 s

first framelast frame

첫 프레임 · 마지막 프레임 (마지막 프레임 = v2 goal). referral 예: wooden dresser with four drawers, chestofdrawers, chest of drawers, wood dresser

4 Place — molmo-augment place (goal-image place 증강)

NextTo 원본에서 preplace 이후 place 구간만 잘라내고 주변 물체를 증강한 클립. 지시문 "Place the X next to the Y". goal_im 3뷰(제3시점, v1 학습에 사용)와 마지막 프레임(v2 학습의 goal)이 있음.

규모: 15,043편(변환 14,567 train + 460 val) · 4,819집 · 101 shard · 121 GB · 클립 길이 중앙값 6.5 s

house_1112 — "Place the yellow kitchen knife next to the glossy black cooking pot with handles" 103 steps ≈ 6.8 s

first framelast framegoal images

첫 프레임 · 마지막 프레임 · goal_im 3뷰 (마지막 프레임 = v2 goal). episode house_1112_p0__trajectories_batch_1_of_2…

house_1091 — "Place the small gray smooth cup without handle next to the modern black toaster" 120 steps ≈ 7.9 s

first framelast framegoal images

첫 프레임 · 마지막 프레임 · goal_im 3뷰 (마지막 프레임 = v2 goal). episode house_1091_p0__trajectories_batch_14_of_…

house_10302 — "Place the plain white ceramic mug with handle next to the teal hardcover book with colorful spine" 79 steps ≈ 5.2 s

first framelast framegoal images

첫 프레임 · 마지막 프레임 · goal_im 3뷰 (마지막 프레임 = v2 goal). episode house_10302_p0__trajectories_batch_19_of…

house_10384 — "Place the food turner next to the soap dispenser" 131 steps ≈ 8.6 s

first framelast framegoal images

첫 프레임 · 마지막 프레임 · goal_im 3뷰 (마지막 프레임 = v2 goal). episode house_10384_p0__trajectories_batch_8_of_…