Skip to content
← All builds
Collect robot training data with a handheld UMI gripper: media published by real-stanford/universal_manipulation_interface
Collect robot training data with a handheld UMI gripper: media published by real-stanford/universal_manipulation_interfaceCredit: real-stanford/universal_manipulation_interface
Collect robot training data with a handheld UMI gripper: media published by real-stanford/universal_manipulation_interface (image 2)
Collect robot training data with a handheld UMI gripper: media published by real-stanford/universal_manipulation_interface (image 3)

Collect robot training data with a handheld UMI gripper

Arm, Hard

Universal Manipulation Interface (UMI) lets you record manipulation demos by hand with a 3D-printed gripper and a GoPro, then train a Diffusion Policy that runs on real arms.

View build on GitHub ↗Share on X

No robot needed to collect data: SLAM on the GoPro video recovers the gripper trajectory. Then deploy on a UR5 or Franka.

Hardware & cost: GoPro, printed gripper, and an arm for deployment; see the hardware guide.

Difficulty: Hard.

Agent install prompt (paste into Claude Code, Cursor or Codex): "Read https://github.com/real-stanford/universal_manipulation_interface, set up its conda env and docker SLAM pipeline, and process the example dataset end to end."

Source: real-stanford/universal_manipulation_interface on GitHub — https://github.com/real-stanford/universal_manipulation_interface (facts checked against the project README, Oct 2026; prices vary by region).

Quick answers

What do you need to collect robot training data with a handheld UMI gripper?
GoPro, printed gripper, and an arm for deployment; see the hardware guide.
How hard is it to collect robot training data with a handheld UMI gripper?
Hard.

How your agent drives it

Copied from the real-stanford/universal_manipulation_interface README

  1. Shell
    $ python run_slam_pipeline.py example_demo_session
  2. Shell
    $ python train.py --config-name=train_diffusion_unet_timm_umi_workspace task.dataset_path=example_demo_session/dataset.zarr.zip

Any coding agent that can run a shell can run these. Read the README before you run anything on a robot that moves.

Ideas to try with it

Our ideas, not the builder's claims.

  1. Record cup-arranging demos by hand with a GoPro, no robot needed
  2. Collect demos in a real kitchen and train a policy back at the lab
  3. Run the example SLAM pipeline on the sample session
1 more idea
  1. Swap the training data for your own task and compare results
Source
real-stanford/universal_manipulation_interface on GitHub
Last commit
2026-05-26
Builder
real-stanford
Repo
real-stanford/universal_manipulation_interface, 1,628 stars
Built with
UR5, Franka

More like this: Arm manipulation

What to build next

Builds that share a board, robot or agent with this one.