InHandRotation_OXT / README.md
Kitval's picture
Upload folder using huggingface_hub
38ef478 verified
|
Raw History Blame Contribute Delete
3.31 kB
metadata
pretty_name: InHandRotation_OXT
task_categories:
  - robotics
tags:
  - tactile
  - visuotactile
  - dexterous-manipulation
  - in-hand-rotation
  - open-x-tactile
size_categories:
  - 1K<n<10K

InHandRotation_OXT

Real-world rollouts of sim-to-real RL in-hand rotation (finger-gaiting) policies on a custom palm-up 4-finger hand built from LEAP Hand finger modules (16 DoF), with TacTip-family marker-based optical tactile sensors on all four fingertips. 10 household objects x 2 policy variants (proprioception + tactile depth/flow, proprioception + tactile depth), 30 s per episode at 10 Hz. Each step contains a fixed external RGB view, 16-D joint positions and the four raw tactile camera images.

Part of the Open-X-Tactile (OXT) collection. Data format (FTP-1 zarr) and key definitions: https://open-x-tactile.github.io/

  • InHandRotation_OXT/: 20 episodes, one zip per episode (episode_<UTC start>.zarr.zip), 6000 steps at 10 Hz.
  • Keys: timestamps, camera_main_rgb (224x224), right_hand_joints (16, rad), right_hand_joints_idx, right_tactile_{data,area,sensor,type}_fingertips (4 x 224x224x3 raw tactile images, sensor TacTipFingertip, type image), sub_task_instruction.
  • Instruction: "Rotate the object in-hand along the palm direction, counterclockwise as viewed from above the palm."
  • Policy variants: full = proprioception + tactile depth + tactile flow; depth = proprioception + tactile depth.
  • Paper: TactileLab: Sim-to-Real Dexterous Robot Manipulation with Efficient Shear-Sensitive Tactile Simulation
zarr object policy steps note
episode_20260627T231758Z.zarr apple full 300
episode_20260627T223535Z.zarr apple depth 300
episode_20260627T232053Z.zarr blue_cylinder full 300
episode_20260627T224323Z.zarr blue_cylinder depth 300
episode_20260627T231028Z.zarr brain full 300
episode_20260627T223322Z.zarr brain depth 300
episode_20260627T225658Z.zarr can full 300
episode_20260627T224205Z.zarr can depth 300
episode_20260627T231333Z.zarr container full 300
episode_20260627T221950Z.zarr container depth 300
episode_20260627T230318Z.zarr gum full 300 object wedged on the fingertips after ~22.5 s; the hand barely moves until the end
episode_20260627T223200Z.zarr gum depth 300
episode_20260627T231502Z.zarr peach full 300
episode_20260627T223916Z.zarr peach depth 300
episode_20260627T230843Z.zarr pilljar full 300
episode_20260627T221542Z.zarr pilljar depth 300
episode_20260627T231159Z.zarr potbox full 300
episode_20260627T224044Z.zarr potbox depth 300
episode_20260627T231632Z.zarr raspberry full 300
episode_20260627T223026Z.zarr raspberry depth 300

Citation

@article{lin2026tactilelab,
  title={TactileLab: Sim-to-Real Dexterous Robot Manipulation with Efficient Shear-Sensitive Tactile Simulation},
  author={Lin, Yijiong and He, Wenhao and Jack, Toby and Yang, Max and Li, Haoran and Lepora, Nathan F.},
  journal={arXiv preprint arXiv:2611.07881},
  year={2026}
}