Datasets:

dex-robot
/

cap

license: other
task_categories:
  - robotics
tags:
  - robotics
  - dexterous-manipulation
  - teleoperation
  - motion-capture
  - human-hand
  - imitation-learning
  - hdf5
size_categories:
  - 10GB

CAP — Dexterous Manipulation Capture Dataset (Cups Ordering, Right Hand)

CAP is a human hand motion capture + dexterous manipulation demonstration dataset. It records an operator performing the cups ordering (cups_order) task with the right hand. Data is captured with a DexCap-style pipeline and ships both compressed trajectory HDF5 files and raw multi-view camera frames, ready for motion retargeting, imitation learning, and hand-pose analysis.

Repo: dex-robot/cap
Main file: cap.zip (10,697,737,890 bytes ≈ 9.96 GiB)
Contents: 18,939 files + 24 directory entries
SHA256: f91c11ee1761822fa875d4783c1fd9388f6fa3f0fa3feeef6cd5c20775220c10
Capture date: 2026-07-07
Task: cups ordering
Hand: right

1. Overview

Item	Value
Task	`cups_order` (ordering cups)
Hand	Right
Capture date	2026-07-07
Recordings	5 (16:46:34 / 16:54:35 / 16:56:54 / 16:59:04 / 17:01:24)
Trajectory files	5 × `.hdf5` (one per recording)
Camera views	6 (left hand / right hand / head stereo / head global RGB / head global depth)
Image frames	18,914 × `.png` + 5 × `.mp4`
Other assets	3 × `.json`, 6 × `.txt`, 5 × `.npy`, 1 × `.raw`
Archive size	10.7 GB (10,697,737,890 bytes)

2. Directory Layout

cap/
├── hdf5/                                   # one compressed trajectory per recording
│   ├── cups_order_right_hand_20260707_164634/
│   │   └── cups_order_right_hand_20260707_164634.hdf5
│   ├── cups_order_right_hand_20260707_165435/
│   │   └── cups_order_right_hand_20260707_165435.hdf5
│   ├── cups_order_right_hand_20260707_165654/
│   │   └── cups_order_right_hand_20260707_165654.hdf5
│   ├── cups_order_right_hand_20260707_165904/
│   │   └── cups_order_right_hand_20260707_165904.hdf5
│   └── cups_order_right_hand_20260707_170124/
│       └── cups_order_right_hand_20260707_170124.hdf5
└── raw_data/                               # raw multimodal stream (images + video + calibration)
    └── cups_order_right_hand_20260707_164634/
        └── cups_order_right_hand_20260707_164634_extracted/
            └── observe/
                └── vision/
                    ├── left_hand_rgb/            # left wrist view (video + metadata)
                    ├── right_hand_rgb/           # right wrist view
                    ├── head_zed_left/            # head ZED left
                    ├── head_zed_right/           # head ZED right
                    ├── head_global_rgb/          # head global RGB
                    ├── head_global_depth/        # head global depth
                    └── *_frames/                 # extracted PNG sequences per view

3. Recording Sessions

Session ID	Start time	HDF5 trajectory	Raw frames
`cups_order_right_hand_20260707_164634`	16:46:34	✅	✅
`cups_order_right_hand_20260707_165435`	16:54:35	✅	—
`cups_order_right_hand_20260707_165654`	16:56:54	✅	—
`cups_order_right_hand_20260707_165904`	16:59:04	✅	—
`cups_order_right_hand_20260707_170124`	17:01:24	✅	—

Note: the full multi-view raw frames (18,915 PNG + 5 MP4) are published with the first session 164634; the other four recordings ship as compressed HDF5 trajectories.

4. Camera Views & Frame Counts

View directory	PNG frames	Video
`head_global_depth_frames`	5,565	✅
`right_hand_rgb_frames`	3,089	✅
`head_zed_right_frames`	2,629	✅
`left_hand_rgb_frames`	2,563	✅
`head_zed_left_frames`	2,546	✅
`head_global_rgb_frames`	2,522	✅

Frames are named frame_%06d.png, but the indices are not strictly contiguous (frames were dropped during capture). Always use the sorted filename order; do not assume i and i+1 are adjacent.

Preview

Preview file	Content
`preview_overview.png`	single-frame overview of all 6 views
`preview_grid.gif`	all 6 views in sync (24 frames)
`preview_right_hand_rgb.gif`	right wrist view
`preview_head_global_rgb.gif`	head global RGB
`preview_head_global_depth.gif`	head global depth

5. Download

5.1 huggingface-cli (recommended, resumable)

pip install -U "huggingface_hub[cli]"
export HF_ENDPOINT=https://hf-mirror.com        # optional mirror for CN networks
huggingface-cli download dex-robot/cap cap.zip --repo-type dataset --local-dir ./cap_data

5.2 Python

from huggingface_hub import hf_hub_download
p = hf_hub_download("dex-robot/cap", "cap.zip", repo_type="dataset")
print(p)

5.3 Direct URL

https://huggingface.co/datasets/dex-robot/cap/resolve/main/cap.zip

6. Extract & Verify

# 1) size
stat -c %s cap.zip        # expected: 10697737890

# 2) SHA256
sha256sum cap.zip
# f91c11ee1761822fa875d4783c1fd9388f6fa3f0fa3feeef6cd5c20775220c10

# 3) CRC check (all 18963 entries must report OK)
unzip -t cap.zip

# 4) extract -> cap/
unzip -q cap.zip

du -sh cap

Expected:

cap/
├── hdf5/      (5 .hdf5 files)
└── raw_data/  (18,934 files)

7. Usage

7.1 Read an HDF5 trajectory

import h5py

path = "cap/hdf5/cups_order_right_hand_20260707_164634/cups_order_right_hand_20260707_164634.hdf5"
with h5py.File(path, "r") as f:
    def show(name, obj):
        if isinstance(obj, h5py.Dataset):
            print(f"D {name:50s} shape={obj.shape} dtype={obj.dtype}")
        else:
            print(f"G {name}")
    f.visititems(show)

7.2 Read image frames

import cv2, glob

frames = sorted(glob.glob(
    "cap/raw_data/cups_order_right_hand_20260707_164634/"
    "cups_order_right_hand_20260707_164634_extracted/observe/vision/"
    "head_global_rgb_frames/frame_*.png"))
print("frames:", len(frames))
img = cv2.imread(frames[0])
print(img.shape)  # (H, W, 3)

8. Recommended Environment

pip install h5py opencv-python numpy

Python ≥ 3.8
~15 GB free disk after extraction (raw PNG sequences are large)
head_global_depth may be 16-bit PNG or accompanied by .npy / .raw raw depth

9. License & Citation

License: see the repo LICENSE; contact the publisher if unspecified.

@misc{cap_dataset_2026,
  title  = {CAP: Cups Ordering Dexterous Manipulation Capture Dataset (Right Hand)},
  author = {dex-robot},
  year   = {2026},
  howpublished = {\url{https://huggingface.co/datasets/dex-robot/cap}}
}

10. FAQ

Q: Download interrupted? huggingface-cli download resumes automatically — just re-run the same command. Add --max-workers 1 for a more stable (slower) transfer.

Q: unzip says "End-of-central-directory signature not found"? The download is incomplete / truncated. Check that the size is exactly 10,697,737,890 bytes and the SHA256 matches; if not, re-download.

Q: I only need trajectories, not the 10 GB of images? Currently only the full bundle is published. Open a Discussion if you need a split release.

Q: How do I align depth with RGB? *_frames and the MP4 of the same view share frame indices (frame_%06d). Cross-view alignment: use the calibration JSON under raw_data/.../observe/vision/.

11. Known Limitations

Frame indices are not contiguous. Each view's frame_%06d.png has gaps (frames dropped during capture) — e.g. right_hand_rgb_frames holds 3,089 frames with a max index of 3,746. Use sorted filename order; never infer the next frame as i+1.
Raw frames cover the first session only. Of the 5 recordings, only 164634 ships full multi-view PNG / MP4; the other four provide HDF5 trajectories only.
HDF5 internal layout may evolve across capture versions. Group / dataset names can differ between recordings — inspect the structure with visititems before reading arrays.
The archive was recovered from an interrupted packing process and passes CRC verification for all 18,963 entries; one incomplete trailing PNG frame was dropped.