metadata
language: en
pretty_name: FINGR
tags:
- robotics
- imitation-learning
- lerobot
FINGR
Demonstrations for FINGR: Learning Dexterous Hand Control for Real-World Rubik's Cube Solving.
36 teleoperated sequences, 576 turns, and 25,520 frames at 10 Hz, collected with an xArm7 and a Sharpa Wave hand. LeRobot v3 format includes joint states and commands, five-finger tactile forces and videos, tracked cube geometry, and turn annotations.
The paper uses 435 turns (18,113 frames) for action training and 24,943 frames for future prediction. The code applies both selections to this complete dataset.