DexYCB Hand Grasping Dataset
Provider：NVIDIA Research / University of Washington
DexYCB provides 1,000 human grasp sequences with ~4.6M RGB-D images across 8 views, per-frame MANO hand parameters and 6D object poses, under CC BY-NC.
At a glance
- Readiness
- L1 · Rigid-body ready
- Rationale
- Manipulation L1: 1,000 real human grasp sequences with per-frame MANO hand parameters, 21 3D joints and 6D object poses, enough to reconstruct hand-object meshes for dynamic replay; the subject is human grasping rather than robot-joint-driven assets, hence L1.
- Scale
- 1,000 grasp sequences (10 subjects × 100 trials) with about 582K frames per view, roughly 4.6 million RGB-D images across 8 views, over 20 YCB objects (The DexYCB project page and paper (CVPR 2021) report 1,000 sequences, 10 subjects, 20 YCB objects, 8 cameras and 582K frames per view; the full archive is 119 GB. Checked 2026-10-25.)
- License
- CC BY-NC (non-commercial) — CC BY-NC 4.0: non-commercial use only, with attribution required. The dataset page states this license explicitly, and commercial use needs separate permission.
- Access
- Open
- Formats
- obj · other
- Runtimes
- other
- Physics properties
- Collision geometryJoint structureSensor attributes
- Domain
- Other
- Published
- 2026-10-25
- Updated
- 2026-10-25
Why it is listed
Dexterous manipulation needs hand-object data with correct contact geometry, and DexYCB is the marker-less reference set: ten subjects grasp twenty standard YCB objects in a calibrated rig of eight synchronized RGB-D cameras, with annotation accurate enough to drive simulation directly.
Readiness assessment
Every frame carries MANO hand parameters, 21 3D joints, 2D keypoints, 6D object poses, masks and bounding boxes, and the YCB meshes ship with the dataset, so rigid object URDFs can be generated for a simulation pipeline. The paper also quantifies an "interaction penalty" — object pose accuracy drops from about 56% to 41% once the object is grasped — which is exactly the regime a manipulation model has to survive.
Practical notes
One thousand sequences of about 72 frames each at 30 Hz amount to roughly five hours of unique recording, distributed as per-subject archives from Google Drive; the license is CC BY-NC 4.0, so research use only.
Disclaimer: license and usage terms follow the official source — verify them yourself before use. This site indexes and links only; it does not host or mirror any asset file.
License and usage terms are those of the official source — verify before use. No asset file is hosted or mirrored on this site.