CCF灵巧装配数据集

# CCF Dexterous Assembly Dataset CCF Dexterous Assembly Dataset is a multi-task real-robot dataset for dexterous-hand assembly and industrial manipulation learning, collected on the `Linkerbot` platform. The dataset is released task by task. Each episode provides robot states and actions, three-view RGB videos, depth data, and related information. ## Task Overview | Task | Task Name | Track | Status | Episodes | Size | |---|---|---|---|---:|---:| | `ccf_nut_grabbing` | Nut grasping and classification (single-hand) | Industrial | Released | 205 | ~71 GiB | | `ccf_tube_pick_and_place` | Test-tube pick and place (single-hand) | Scientific | Coming soon | — | — | ## Task Description ### Task 1: `ccf_nut_grabbing` — Nut grasping and classification (single-hand) The right hand picks up a nut from the tabletop and places it into the corresponding compartment of a three-slot storage box according to the nut size. | Item | Description | |---|---| | Objects | Nuts, three-slot storage box | | Nut sizes | `M27` (small), `M33` (medium), `M45` (large) | | Operating hand | Single right-hand operation | ### Task 2: `ccf_tube_pick_and_place` — Test-tube pick and place (single-hand, coming soon) The right hand picks a test tube from a source rack, moves it to a target rack, and places it in an empty slot. | Item | Description | |---|---| | Objects | Test tubes, test-tube racks | | Tube sizes | `50 ml`, `15 ml` | | Operating hand | Single right-hand operation | ## Dataset Overview The statistics below cover only the released task (`ccf_nut_grabbing`): | Item | Value | |---|---:| | Number of tasks | 1 (out of 2 planned) | | Episodes | 205 | | Total collection duration | 1.84 h | | Episode duration | 25.0 s ~ 45.3 s, average 32.3 s | | Frame rate | 30 FPS | | Camera views | Head + left wrist + right wrist (3 in total) | | Data formats | `JSON`, `NPZ`, `NPY`, `MP4`, `TAR` | | Size | ~71 GiB | ## Data Access ### Download :modelscope-code[]{type="sdk"} :modelscope-code[]{type="git"} ### Layout and Extraction Each task is released as **split tar volumes**. Repository layout: ```text ccf_nut_grabbing/ ├── ccf_nut_grabbing_part_a.tar # episode_000000 ~ episode_000100 (101 episodes) ├── ccf_nut_grabbing_part_b.tar # episode_000101 ~ episode_000204 (104 episodes) └── checksums.txt # SHA-256 checksums ``` Extract both tar volumes after downloading: ```bash tar -xf ccf_nut_grabbing_part_a.tar tar -xf ccf_nut_grabbing_part_b.tar ``` You will get two sibling directories `ccf_nut_grabbing_part_a/` and `ccf_nut_grabbing_part_b/`, totaling 205 episodes. Usage examples and tooling are available in the publisher's open-source organization: <https://github.com/linker-bot> ## Hardware and Collection System The data is generated by the `Linker Open TeleDex` collection system. The hardware consists of `Linkerbot` proprietary products: | Module | Model | |---|---| | Robot arm | `Linker Arm A7 Lite` | | Dexterous hand | `Linker Hand O6` | | Teleoperation master arm | `Linker TA` | | Teleoperation glove | `Linker FFG` | ## Directory Structure After extracting both tar volumes you get two sibling directories with non-overlapping episode ranges: ```text ccf_nut_grabbing_part_a/ # episode_000000 ~ episode_000100 ├── episode_000000/ ├── ... └── episode_000100/ ccf_nut_grabbing_part_b/ # episode_000101 ~ episode_000204 ├── episode_000101/ ├── ... └── episode_000204/ ``` Per-episode structure is identical: ```text episode_000000/ ├── telemetry.npz ├── camera_info.json ├── manifest.json ├── metadata.json └── cameras/ ├── cam_top.mp4 ├── cam_top.timestamps.npy ├── cam_top_depth.tar ├── cam_top_depth.timestamps.npy ├── cam_top_depth_colormap.mp4 ├── cam_left_wrist.mp4 ├── cam_left_wrist.timestamps.npy ├── cam_left_wrist_depth.tar ├── cam_left_wrist_depth.timestamps.npy ├── cam_left_wrist_depth_colormap.mp4 ├── cam_right_wrist.mp4 ├── cam_right_wrist.timestamps.npy ├── cam_right_wrist_depth.tar ├── cam_right_wrist_depth.timestamps.npy └── cam_right_wrist_depth_colormap.mp4 ``` ## Episode Data Format | File | Format | Description | |---|---|---| | `telemetry.npz` | NPZ | Aligned robot states, actions, end-effector poses, IMU, and synchronization-validation fields | | `camera_info.json` | JSON | Color / depth / IMU calibration for the three cameras | | `manifest.json` | JSON | Episode index: duration, frame count, frame rate, collection flags, and key file paths | | `metadata.json` | JSON | Conversion settings, alignment settings, synchronization error, video encoding parameters, device and recording summary | | `cameras/*.mp4` | MP4 | Three-view RGB videos | | `cameras/*.timestamps.npy` | NPY | RGB video frame timestamps (seconds) | | `cameras/*_depth.tar` | TAR | Packed depth image sequence | | `cameras/*_depth.timestamps.npy` | NPY | Depth frame timestamps (seconds) | | `cameras/*_depth_colormap.mp4` | MP4 | Depth colormap preview video (for quick inspection only) | ## Core Field Descriptions ### telemetry.npz `telemetry.npz` stores the unified aligned timeline for all modalities, together with the robot state and control actions on that timeline. Fields in the example `episode_000000`: | Field | Shape | dtype | Description | |---|---|---|---| | `timestamps` | `(N,)` | `float64` | Aligned unified timeline (seconds, Unix epoch) | | `qpos` | `(N, 26)` | `float32` | Joint positions | | `qvel` | `(N, 26)` | `float32` | Joint velocities | | `effort` | `(N, 26)` | `float32` | Torque / force / current state values | | `actions` | `(N, 26)` | `float32` | Control action sequence (consistent with teleoperation commands) | | `sync_validation_max_diff` | `(N,)` | `float64` | Maximum cross-sensor time difference per frame (milliseconds) | | `imu_cam_top` / `imu_cam_left_wrist` / `imu_cam_right_wrist` | `(N, 6)` | `float32` | Per-camera aligned IMU readings `[ax, ay, az, gx, gy, gz]` | | `ee_poses_qpos_left` / `ee_poses_qpos_right` | `(N, 7)` | `float32` | End-effector poses from left / right arm state `[x, y, z, qx, qy, qz, qw]` | | `ee_poses_actions_left` / `ee_poses_actions_right` | `(N, 7)` | `float32` | End-effector poses from left / right arm actions | `N` is the number of aligned frames. The 26 dimensions of `qpos / actions` are sliced as follows: | Slice | Meaning | |---|---| | `0:7` | Left arm | | `7:14` | Right arm | | `14:20` | Left hand | | `20:26` | Right hand | ### manifest.json | Field | Type | Description | |---|---|---| | `episode_id` | string | Episode identifier | | `format_version` | string | Format version | | `created_at` | string | Creation time | | `duration` | number | Episode duration (seconds) | | `frame_count` | number | Number of aligned frames | | `fps` / `dt` | number | Target frame rate and frame interval | | `collect_depth` / `collect_tactile` / `collect_imu` | boolean | Modality collection flags | | `sync_error` | object | Synchronization error statistics | | `files` | object | Index of key file paths | ### metadata.json Contains `conversion`, `alignment`, `video`, `ee_pose`, `imu`, `device`, and `recording` fields, recording the conversion pipeline, alignment strategy, video encoding parameters, and device information. ### camera_info.json Top-level keys are `cam_top` / `cam_left_wrist` / `cam_right_wrist`. Each camera records three types of information: `color`, `depth`, and `imu`. ## License This dataset is released under a **dual license**: - **Default license**: [Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International (`CC BY-NC-SA 4.0`)](https://creativecommons.org/licenses/by-nc-sa/4.0/). Permits non-commercial research and educational copying, distribution, display, and modification; derivative works must be released under the same or a compatible license. - **Commercial license**: Commercial use (including training commercial models with this dataset) is not covered by the default license and requires prior written commercial authorization from the publisher. For commercial cooperation, please contact the publisher via the official website. ## Citation ```bibtex @misc{ccf_dexterous_assembly_dataset, title = {CCF Dexterous Assembly Dataset}, author = {{Linkerbot Beijing Technology Co., Ltd.}}, year = {2026}, howpublished = {\url{https://www.modelscope.cn/datasets/Linkerbot/CCF-Dexterous-Assembly-Dataset}}, } ``` Acknowledgment template: > Data source: CCF Dexterous Assembly Dataset, collected by Linkerbot Beijing Technology Co., Ltd. ## Contact - Publisher: Linkerbot Beijing Technology Co., Ltd. - Website: https://linkerbot.cn