SharpaDex v1.0 Download

SharpaDex

300 hours of real bimanual teleoperation. Two 22-DoF hands, touch on every fingertip, every frame.

Move across the words. Every particle is a task, sized by its hours.

Every fingertip, every frame.

Ten vision-based tactile sensors record 6-axis force and deformation images at 30 fps, aligned with four cameras and 65 joints. Below, a real episode of hanging a shirt, replayed from its fingertip force: each line is one fingertip, and it lights up only while that finger is touching.

  1. Left thumb
  2. Left index
  3. Left middle
  4. Left ring
  5. Left little
  6. Right thumb
  7. Right index
  8. Right middle
  9. Right ring
  10. Right little

Moments from 300 hours.

Every card is a real episode, one per task, stamped with the day it was recorded. Move across the stack, or tap it, to bring the next one forward.

Three levels of language, on every episode.

A task description, subtasks with start and end frames, and the atomic action of each subtask. One real episode:

What a frame contains.

Joints
65-D state and action: two 7-DoF arms, two 22-DoF hands, a 7-DoF torso
Force
60-D fingertip wrench; joint torque of both arms and hands
Format
LeRobot v3.0 and v2.1, H.264 video
Licence
CC BY-NC 4.0

Two hands, 44 joints, ten fingertips.

The recorded joint states pose the open-source Sharpa Wave model frame by frame, next to what the head camera saw. A fingertip lights up while it touches, brighter the harder it presses.

22
joints per hand
10
fingertips with 6-axis force
30 fps
joints, force and video in sync

Drag to turn · double-click to reset

59 tasks, from ironing to pipetting.

Sorted by hours of data. Most of it is tool use, cloth and paper, insertion and twisting — the work a gripper can't do. Hover a task to watch it.

    Get the data.

    Download

    Clone one task or all of it. Every season is a LeRobot root.

    git lfs install
    git clone https://huggingface.co/datasets/Sharpa-Robotics/SharpaDex-v1.0

    Watch first

    Browse every episode in the browser — cameras, touch, hands and language in sync.

    Open the viewer

    Cite

    @misc{sharpadex_v1_2026,
      title        = {SharpaDex v1.0: Large-Scale Bimanual Dexterous Manipulation Demonstrations},
      author       = {{Sharpa}},
      howpublished = {\url{https://huggingface.co/datasets/Sharpa-Robotics/SharpaDex-v1.0}},
      year         = {2026}
    }

    Team

    SharpaDex was collected, annotated and released by Sharpa. Questions about the dataset: academia@sharpa.com.

    Acknowledgements

    Thanks to the teleoperators and annotators who recorded and labelled every episode. SharpaDex uses the LeRobot dataset format and is hosted on the Hugging Face Hub. This site is built with three.js and urdf-loader; the hand models are the open-source Sharpa Wave assets.