Datasets:
Dexterous Hand RGB-D Motion Sequences
Fixed-camera, Blender-rendered RGB-D and semantic-label sequences for nine articulated dexterous hands and grippers. Each independent actuator is swept from its lower to upper URDF limit and then gradually back while the other independent actuators remain at their lower limits. Mimic joints follow their URDF definitions. These are synthetic renders, not physical camera recordings.
Contents
Each directory under data/ contains:
rgba/NNNN.png: 800x800 transparent RGBA render.depth/NNNN.png: 16-bit depth in millimetres; zero denotes background.semantic/NNNN.npy: integer label image;-1background,0static geometry, and positive labels map to joints/links viaposes.json.poses.json: joint states, camera calibration, semantic-label mapping, and source URDF path relative to the upstream dex-urdf repository.motion.mp4,motion.gif,contact_sheet.png: visual previews.
manifest.json summarizes hand names, frame counts, actuators, and camera IDs.
Frames are rendered sequentially; the previews play at approximately 4.55 fps.
Provenance and licenses
The hand geometry and kinematic descriptions come from
DexSuite/dex-urdf, with their included
upstream sources and modifications. Per-hand license files and the upstream
citation are in licenses/. The source repository reports different licenses
for different models, and some notices conflict or do not state a license.
Accordingly, license: other is used here and no single license is asserted
for the complete dataset. Review LICENSES.md and the copied notices before
redistributing or reusing any hand's renders.
Please cite the upstream dex-urdf work as described in licenses/dex-urdf/CITATION.cff.
- Downloads last month
- 659