--- license: apache-2.0 task_categories: - robotics tags: - robotics - manipulation - imitation-learning - elevator-button-press - isaac-sim - isaac-lab - omy pretty_name: ev Button Push v02 (OMY call-button, contact-force teacher) size_categories: - 10B 이전 collection (거리 기반 v01) 은 > [`RobotisAI/evButtonPush-260422-01-hdf5`](https://huggingface.co/datasets/RobotisAI/evButtonPush-260422-01-hdf5) > 에서 받을 수 있습니다 (baseline / ablation 용). ### v02 특성 (권장) - **45 demos**, seed `0`~`44`, 전부 pressed=True - 약 35초/ep, ~1050 프레임 @ 30fps - 성공 판정: **물리 접촉 힘 ≥ 0.5 N** (PhysX `tip_contact` sensor) - 엘베 **문 FSM**: `state 0 (대기) → 1 (눌림, LED on) → 2 (문 여는 중) → 3 (완전 개방)` - 평균 접촉 힘 피크: 60~80 N - Press 목표 Z offset: `button_base` 기준 **+0.11 m** 고정 - `call_button_lit` LED 관찰 포함 - Teacher CLI: ``` --approach-steps 100 --press-steps 400 --stall-steps 20 --hold-steps 30 --return-steps 300 --home-idle-steps 150 --press-depth 0.035 --contact-force-threshold 0.5 ``` ### Task - Task ID (Isaac Lab): `RobotisLab-CallButton-Right-OMY-v0` - 로봇: ROBOTIS **OMY** (6축 팔 + 2지 그리퍼) - Pedestal 높이: 0.908 m, 그리퍼 항상 닫힘 (stiffness 2×10⁶, effort 5000) - 엘리베이터 USD: 100종 variants 중 seed별 random_choice (`elevator_setup_new`) - 로봇 초기 자세 랜덤 (seed별): - 버튼까지 거리(standoff): 0.55 ~ 0.65 m - 좌우 lateral: −0.05 ~ 0.05 m - yaw: 0° ### 카메라 3대 모두 **RGB, 1920 × 1200, Pinhole**: - 수평 aperture 20.955 mm - 수직 aperture 13.097 mm (16:10 비율로 자동 계산) - clipping (0.01, 100) m | 이름 | prim_path | 기준 Frame | 위치 (m) | Quat (w, x, y, z) | FOV (H / V / D) | focal | |---|---|---|---|---|---|---| | `cam_wrist` | `Robot/OMY/link6/cam_wrist` | link6 (팔과 함께 이동) | (0.000, −0.080, 0.070) | (0.0018, −0.0018, 0.7071, 0.7071) | 83.21° / 58.06° / 92.64° | 11.8 mm | | `cam_top` | `Robot/OMY/world/cam_top` | 로봇 base | (0.000233, 0.1499, 0.43953) | (0.5144, 0.4146, −0.5064, −0.5542) | 92.67° / 66.44° / 102.03° | 10.0 mm | | `cam_belly` | `Robot/OMY/world/cam_belly` | 로봇 base | (0.08705, 0.1499, 0.1075) | (−0.4901, −0.5097, 0.5097, 0.4901) | 92.67° / 66.44° / 102.03° | 10.0 mm | OMY USD 내부의 `world` prim은 시뮬레이션 월드가 아니라 **로봇 base (link0)**를 가리킴. ``` world_pose(cam_top) = robot.root_pose_w ⊙ offset_cam_top world_pose(cam_wrist) = robot.root_pose_w ⊙ FK(link0→link6) ⊙ offset_cam_wrist ``` ### HDF5 내부 구조 (episode 당) ``` data/demo_0/ ├── actions (T, 7) float32 arm 6 + gripper 1 ├── processed_actions (T, 10) float32 전체 joint target ├── obs/ │ ├── cam_wrist (T, 1200, 1920, 3) uint8 │ ├── cam_top (T, 1200, 1920, 3) uint8 │ ├── cam_belly (T, 1200, 1920, 3) uint8 │ ├── joint_pos (T, 10) float32 │ ├── joint_vel (T, 10) float32 │ ├── call_button_pos (T, 3) float32 버튼 world 좌표 │ ├── rel_ee_call_button_distance (T, 3) float32 tip-button 벡터 │ └── call_button_lit (T, 1) bool ⭐ LED on/off ├── states/articulation/robot/... ├── states/rigid_object/pedestal/... └── initial_state/... ``` ### 파일 구성 - `demo_XX.hdf5` - episode 하나당 파일 하나 - `demo_XX.log` - teacher stdout (`[SETUP]`, `[PHASE]`, `[PRESSED]`, `[RESULT]` 마커) - `summary.log` - demo당 한 줄 요약 ### 추천 사용 방법 Imitation learning (ACT, Diffusion Policy, BC) 학습 시 `qpos_dim = 8`: ```python qpos = concat(joint_pos[:6], gripper_closed_flag, call_button_lit) # 8 action = [arm_delta[:6], gripper_cmd] # 7 images = {cam_wrist, cam_top, cam_belly} # 1920×1200 (필요시 resize) ``` 일반적으로 학습 파이프라인에서는 **240×320으로 resize**해서 씁니다. 원본은 1920×1200으로 보존. --- ## English Simulated demonstrations of a **ROBOTIS OMY** 6-DoF arm pressing an elevator call-button, collected in **NVIDIA Isaac Sim / Isaac Lab** with a scripted IK teacher. This dataset is the **v02** collection — improved over v01 with physics-based success detection and elevator-door FSM. > The earlier v01 collection (distance-based teacher, baseline only) is at > [`RobotisAI/evButtonPush-260422-01-hdf5`](https://huggingface.co/datasets/RobotisAI/evButtonPush-260422-01-hdf5). ### v02 highlights (recommended for training) - **45 demos**, seeds `0` – `44`, all pressed = True - ~35 s per episode, ~1050 frames at 30 fps - Success criterion: **physical contact force ≥ 0.5 N** (PhysX `tip_contact` sensor) - Elevator **door FSM** progresses on press: `state 0 (idle) → 1 (pressed / LED on) → 2 (door opening, progress 0→1) → 3 (open)` - Average peak gripper tip contact force: 60 – 80 N - Press target Z offset: `+0.11 m` relative to `button_base` (fixed inside teacher). - Includes `call_button_lit` LED observation. - Teacher CLI: ``` --approach-steps 100 --press-steps 400 --stall-steps 20 --hold-steps 30 --return-steps 300 --home-idle-steps 150 --press-depth 0.035 --contact-force-threshold 0.5 ``` ### Task - Task ID (Isaac Lab): `RobotisLab-CallButton-Right-OMY-v0` - Robot: ROBOTIS **OMY** (6-DoF arm + 2-finger gripper) - Pedestal: 0.908 m, gripper always closed (stiffness 2 × 10⁶, effort 5000) - Elevator USD: random choice from 100 variants (`elevator_setup_new`) - Robot initial pose randomization (per-seed): - standoff distance to button: 0.55 ~ 0.65 m - lateral offset: −0.05 ~ 0.05 m - yaw: 0° ### Cameras All three cameras record **RGB, 1920 × 1200, Pinhole**: - horizontal aperture 20.955 mm - vertical aperture 13.097 mm (derived from 16:10 aspect) - clipping (0.01, 100) m | Name | Prim path | Frame | Position (m) | Quat (w, x, y, z) | FOV (H / V / D) | focal | |---|---|---|---|---|---|---| | `cam_wrist` | `Robot/OMY/link6/cam_wrist` | link6 (local, moves with arm) | (0.000, −0.080, 0.070) | (0.0018, −0.0018, 0.7071, 0.7071) | 83.21° / 58.06° / 92.64° | 11.8 mm | | `cam_top` | `Robot/OMY/world/cam_top` | robot base | (0.000233, 0.1499, 0.43953) | (0.5144, 0.4146, −0.5064, −0.5542) | 92.67° / 66.44° / 102.03° | 10.0 mm | | `cam_belly` | `Robot/OMY/world/cam_belly` | robot base | (0.08705, 0.1499, 0.1075) | (−0.4901, −0.5097, 0.5097, 0.4901) | 92.67° / 66.44° / 102.03° | 10.0 mm | The `world` prim inside the OMY USD refers to the robot's base frame (link0), **not** the simulation world. ``` world_pose(cam_top) = robot.root_pose_w ⊙ offset_cam_top world_pose(cam_wrist) = robot.root_pose_w ⊙ FK(link0→link6) ⊙ offset_cam_wrist ``` ### HDF5 layout per episode ``` data/demo_0/ ├── actions (T, 7) float32 ├── processed_actions (T, 10) float32 ├── obs/ │ ├── cam_wrist (T, 1200, 1920, 3) uint8 │ ├── cam_top (T, 1200, 1920, 3) uint8 │ ├── cam_belly (T, 1200, 1920, 3) uint8 │ ├── joint_pos (T, 10) float32 │ ├── joint_vel (T, 10) float32 │ ├── call_button_pos (T, 3) float32 │ ├── rel_ee_call_button_distance (T, 3) float32 │ └── call_button_lit (T, 1) bool ⭐ LED state ├── states/articulation/robot/... ├── states/rigid_object/pedestal/... └── initial_state/... ``` ### Per-version files - `demo_XX.hdf5` — one episode per file - `demo_XX.log` — scripted-teacher stdout with `[SETUP]`, `[PHASE]`, `[PRESSED]`, `[RESULT]` markers - `summary.log` — one line per demo ### Suggested usage Imitation learning (ACT, Diffusion Policy, BC) with `qpos_dim = 8`: ```python qpos = concat(joint_pos[:6], gripper_closed_flag, call_button_lit) # 8 action = [arm_delta[:6], gripper_cmd] # 7 images = {cam_wrist, cam_top, cam_belly} # 1920×1200 (resize as needed) ``` For efficiency, training pipelines commonly resize to 240 × 320 before consumption. We retain the original 1920 × 1200 to keep the data lossless.