YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
DexT2 metadata and example package (937-episode release)
This package corresponds to the current Hugging Face release airis-lab-ufl/DexT2.
- Tasks: 32
- Trajectories: 937
- Frames: 407,867
- Original recording rate: 25 Hz
- Active trajectory duration: 4.53 h (271.9 min), computed as 407,867 / 25 / 3,600.
Note: The older release included sess_20261013_201653_episode_0017.zarr, sess_20261021_231911_episode_0012.zarr and sess_20261021_233934_episode_0034.zarr. That episode is not in the current HF release.
Precheck and VLM review
All 937 / 937 current trajectories pass structural and metadata precheck with 0 errors and 0 warnings. It contains 97 fail findings across 68 episodes.
We manually verified all the 97 fail, and found that: except for manipulation of lightweight objects (fabric and zip tie), the noisy tactile signal can be mostly reduced by adopting a simple baseline: A tactile sensor reading caused by contact is obtained as a residual sensor reading by subtracting the pre-contact noisy baseline reading. However, this baseline still requires manual verification to determine the timestamp of the pre-contact baseline tactile sensor reading by observing the wrist camera view.
Limitations
For tactile gloves such as Wuji Gloves, the sensor reading can be affected by the hand motion due to the stretch and compression on the fabric as well as the taxels. Also, the adopted baseline still requires manual verification on the timestamp of the pre-contact tactile reading. Therefore, we suggest better approach should be developed to improve the noisy sensor reading. Also, please be awared that for manipulation of lightweight object(In our case, task includes fabric and zip tie), our baseline may not seperate the residual tactile reading caused by the light contact out of the noisy raw tactile reading.
- Downloads last month
- 460