handumi-towel-piper-562-640

HandUMI towel folding: retargeted to bimanual AgileX PiPER (640×480)

The raw capture (murobotics/handumi-towel-raw-562) retargeted by inverse kinematics to bimanual AgileX PiPER: 14-D joint states and actions (radians, grippers in [0, 1]) with the wrist views resized to 640×480. 562 episodes, 200,835 frames. This is the training set of murobotics/handumi-towel-act-562.

Recorded with HandUMI, a hand-worn bimanual interface whose swappable fingertips match the robot's installed parallel-jaw gripper. Task: grasp two corners of a flat yellow towel and fold it in half. 562 episodes, 30 fps, 111.9 minutes, no robot-collected data. Created with LeRobot (v3 format).

• Project website: https://murobotics.ai/research/handumi
• Hardware: https://github.com/murobotics-ai/handumi-hw
• Software: https://github.com/murobotics-ai/handumi-sw
• Documentation: https://murobotics-ai.github.io/handumi-sw/

Collected with the NONHUMAN community. Originally published as NONHUMAN-RESEARCH/yellow-towel-562-piper-640.

License

CC BY 4.0.

Citation

@misc{meza2026handumi,
  title  = {{HandUMI}: A Modular Hand-Worn Parallel-Jaw Interface for
            Robot-Free Bimanual Data Capture},
  author = {Meza, Brik and Perez, Leonardo and Mendoza, Alvaro and
            Ubillus, Aniela and Ramos, Mitshell and Bastidas, Raul and
            Olivera, Mauricio and Bastidas, Bryan},
  year   = {2026},
  url    = {https://murobotics.ai/research/handumi}
}