Datasets:
You need to agree to share your contact information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset is shared with collaborators and reviewers on request. Please state your name, affiliation, and intended use. Access is granted manually.
Log in or Sign Up to review the conditions and access this dataset content.
PST MugBox-v2 Teleoperation Demonstrations
UR5e + Inspire Hand(右手・6 能動関節)による MugBox-v2 タスクの遠隔操作デモ集です。
Leap Motion で捉えた人の手を Inspire Hand へリターゲティングし、手首姿勢は Pinocchio CLIK
(IK)で UR5e に追従させて ManiSkill / SAPIEN 上で収集しました。成功試行のみ(success=True)
を保存しています。
研究の主題は「非専門家オペレータから高品質なデモを大量に集めるにはどうすればよいか」です。 収集時にオペレータへ提示する視覚表現(ロボットハンド上の点群アノテーション)を 3 条件に分け、 専門家データと比較できるように構成されています。
アクセス
public + gated(手動承認) です。HF 上でアクセス申請をしてください。承認後にダウンロードできます。
hf auth login
hf download yamaama/pst-demo --repo-type dataset \
--include "inspire_MugBox-v2/expert/ymgc/*" --local-dir ./pst-demo
収集条件
オペレータへの提示内容だけが異なり、タスク・ロボット・シーンは共通です。
条件の定義はコードリポジトリの configs/keypoint/*.json に対応します。
| 条件 | オペレータへの提示 | 対応する keypoint config |
|---|---|---|
baseline |
点群アノテーションなし(ロボットのみ) | — |
inspire_standard |
全 5 指を色分けした点群を表示 | inspire_standard.json(keypoint_indices=[0,1,2,3,4]) |
inspire_salient |
親指・人差し指のみリッチ表示、他 3 指はグレーで接触フィードバックなし | inspire_salient.json(keypoint_indices=[0,1]) |
expert |
熟練オペレータによる収集(参照上限) | — |
各エピソードの HDF5 属性 operator_saw_particles に、収集時に点群が表示されていたかが記録されています。
構成
inspire_MugBox-v2/
├── baseline/ 05/ 08/ 09/ 14/ 20/ 21/
├── inspire_standard/ 06/ 10/ 12/ 18/ 19/ 30/
├── inspire_salient/ 07/ 15/ 25/ 27/ 31/ 32/
└── expert/ 01/ 22/ 23/ ymgc/
数字はオペレータ ID(匿名化済み)で、その直下に episode_<time>_<date>.hdf5 が並びます。
| 条件 | オペレータ数 | エピソード数 | 容量 |
|---|---|---|---|
baseline |
6 | 320 | 約 13 GB |
inspire_standard |
6 | 352 | 約 15 GB |
inspire_salient |
6 | 442 | 約 19 GB |
expert |
4 | 360 | 約 14 GB |
| 合計 | 22 | 1,474 | 約 61 GB |
以下の派生データは 含めていません(コードから再生成できるか、生デモへのシンボリックリンクのため)。
*_contact… 接触力を再計算した非破壊コピー(src/osx/data_logger/recompute_torques.py)*_notip… 指先点群を除いた派生版_mixed_N100/_mixedN100_camfix/_realparam… 条件混合サンプリング(生デモへのリンク集)
HDF5 スキーマ
T はエピソード長(可変、概ね 200〜400 ステップ)。画像・深度はすべて 224×224、gzip 圧縮。
observations/
| キー | 形状 | dtype | 内容 |
|---|---|---|---|
images_robot_only |
(T, 224, 224, 3) | uint8 | メインカメラ、ロボットのみ |
images_cam_top_robot |
(T, 224, 224, 3) | uint8 | 俯瞰カメラ |
images_cam_side_robot |
(T, 224, 224, 3) | uint8 | 側方カメラ |
images_cam_forward_robot |
(T, 224, 224, 3) | uint8 | 前方カメラ |
images_cam_hand_robot |
(T, 224, 224, 3) | uint8 | 手先カメラ |
images_operator_robot |
(T, 224, 224, 3) | uint8 | 収集時にオペレータが見ていた視点 |
depth, depth_cam_top, depth_cam_side, depth_cam_forward, depth_cam_hand, depth_operator |
(T, 224, 224, 1) | float16 | 各カメラのカメラ座標系 Z [mm](負値) |
joint_positions |
(T, 13) | float32 | ee_pos(3) + ee_quat_xyzw(4) + finger_qpos(6) |
object_poses |
(T, 7) | float32 | 操作対象の pose pos(3) + quat_wxyz(4) |
その他
| キー | 形状 | 内容 |
|---|---|---|
actions |
(T, 13) | target_ee_pos(3) + target_ee_quat_xyzw(4) + finger_active(6) |
timestamps |
(T,) | 収集時刻 |
success |
() | エピソードの成否(本データセットは全て True) |
analysis/joint_torques |
(T, 6) | 解析用トルク |
diagnostics/wrist_pos, wrist_rot |
(T,3), (T,4) | Leap Motion 手首位置・姿勢(スムージング後) |
diagnostics/retarg_qpos |
(T, 12) | dex-retargeting 出力 |
diagnostics/target_ee_pos, cur_ee_pos |
(T, 3) | IK 目標 EE 位置・実際の EE 位置 |
diagnostics/arm_delta_pos, arm_delta_rot |
(T, 3) | 送信した位置・回転デルタ |
diagnostics/arm_qpos |
(T, 6) | UR5e 関節角 [rad](SAPIEN DOF 順) |
diagnostics/qvel, qf |
(T, 18) | 関節速度・トルク(アーム 6 + ハンド 12) |
diagnostics/contact_f_wrist, contact_f_hand |
(T, 3) | wrist_3_link / hand_base_link への接触力 |
属性(attrs)
env_id(MugBox-v2)、robot_uids(ur5e_inspire_hand_right)、arm_control(ik_pose)、
hand_dof=6、arm_dof=7、episode_length、sim_seed、success、operator_saw_particles、
object_initial_pos(3) / object_initial_quat(4, wxyz)、
image_height / image_width / camera_fovy / camera_pose_p(3) / camera_pose_q(4, wxyz)。
行動空間は 13 次元(EE 位置 3 + クォータニオン 4 + Inspire Hand 能動関節 6)。EE 目標は Pinocchio CLIK で UR5e の関節角に変換されてから送信されます。
読み込み例
import h5py
with h5py.File("inspire_MugBox-v2/expert/ymgc/episode_XXXXXX_XXXXXXXX.hdf5") as f:
actions = f["actions"][:] # (T, 13)
qpos = f["observations/joint_positions"][:] # (T, 13)
rgb = f["observations/images_robot_only"][:]
print(dict(f.attrs))
点群ベースの学習(DP3)に使う場合は、コードリポジトリの
src/mani_skill/convert_osx_to_dp3.py で深度画像から点群へ変換してください。
ライセンス
未確定です(共著者間での共有を目的とした gated 公開)。外部公開時に確定します。
- Downloads last month
- 12