Open Source RoboticsRobots, software & learning

UMI

Robot / hardware build

Handheld gripper-based demonstration collection and diffusion-policy workflow, with separate robot deployment hardware and explicit SLAM, timing and source-artifact gaps. [1] [5] [6] [7]

Build & first use →Official website

For
Researchers able to assemble capture hardware and validate robot-specific policy deployment [1]
Cost
No current cost or bill of materials established yet.
Source rights
Some licenses are unresolved. Review licenses
Unknowns
tools not fully established. Review gaps

Source researched · Not tested by us
Sources checked · Research details

Upstream project illustration showing UMI data collection and deployment examples; the host robots are separate equipment.
Upstream project illustration showing UMI data collection and deployment examples; the host robots are separate equipment. Columbia Artificial Intelligence and Robotics Lab and UMI contributors · MIT. Original image unchanged; displayed responsively. Preview re-encoded for faster display; the link opens the original file.
Licenses & related hardware
License
MIT · license not fully assessed for mechanical design · Details by part

Requirements · Research details and dates

On this page

Overview

Universal Manipulation Interface uses a handheld parallel-jaw gripper with a camera to collect demonstrations, then processes those observations into policy-training data. The selected repository revision is d095ba9590df789df5189eea5ee7e431689038a6. The handheld capture device and the robot deployment system are distinct builds with different costs and dependencies. [1] [2] [5] [6]

Is this for you?

Known limitations
  • SLAM robustness limits usable demonstrations
  • Cup-policy README reports poor direct-sunlight behavior
  • External CAD and dataset/checkpoint payloads were not preserved here
  • Example checkpoint download and evaluation filenames differ

[1] [5] [9]

What you need

Operating system
  • Ubuntu 22.04 is the only tested installation environment reported in the selected README

[1]

Host computer

Training/inference workstation; README reports single-GPU training tested on RTX3090 24GB, not a proven universal minimum.

[1]

Dependencies
  • Capture gripper, GoPro/lens and storage
  • Separate ORB_SLAM3 fork and Docker pipeline
  • Training data/checkpoint and diffusion-policy environment
  • Deployment robot, WSG50, camera capture chain and workstation for selected UR route

[1] [5] [7]

Hardware & parts

Hardware configuration

Handheld parallel-jaw data-collection gripper; robot-mounted deployment gripper and arm are separate hardware.

[5] [6]

Sensors
  • GoPro camera with lens accessory for selected capture route
  • Visual markers and camera/IMU-based reconstruction pipeline

[1] [5] [6]

Assembly & first use

Project documentation [1]

Calibration

Match robot/gripper addresses, observation/action latency, TCP offset and geometry settings to the actual deployment; example values are not measurements of another rig.

[7]

Cost

No current, complete configuration-specific build cost has been established for this profile.

3D preview

Preview not available yet. Design-source links are available, but the reviewed record does not establish unambiguous reuse rights for a hosted CAD preview.

Project notes

The hardware guide describes two packages

The capture guide lists printed gripper parts, linear rails, springs, soft fingers, visual tags and a GoPro Hero9 with a Max Lens Mod. It links separate Onshape documents for the gripper and fingers. It also states that left and right grippers use different tags, so bimanual hardware cannot be treated as two unconfigured copies. The public text export was preserved; image-based assembly details and external CAD payloads were not independently inspected. [5]

The deployment section instead lists a UR5-family arm, Weiss WSG-50, camera accessories, an HDMI capture card and workstation connections. Its per-component historical prices do not establish a current complete delivered budget, and the robot is listed by quote. A low capture-device cost therefore should not be presented as the cost of a complete autonomous robot system. [5]

SLAM quality limits the training input

The README provides an example SLAM workflow and explicitly identifies its ORB_SLAM3 fork as a fragile stage. Its example success fraction belongs to that supplied session, not to every new recording. The fork, container and data-collection instructions are separate dependencies. No videos were processed and no SLAM result was reproduced here. [1]

The selected installation instructions report testing only on Ubuntu 22.04 and name an RTX3090 with 24 GB for a tested single-GPU training setup. These are reported environment details, not a catalog-tested minimum specification. The workflow links processed datasets and pretrained checkpoints externally; their complete payloads and applicable licenses were not preserved in this review. [1] [9]

Deployment still requires rig-specific configuration

The example configuration contains robot and gripper addresses, observation and action latencies, a TCP offset, collision-geometry parameters and an inter-arm transform. Those are concrete reasons to match a deployment to the physical rig rather than treating broad calibration-free project language as eliminating all setup. They are examples, not measurements of a reader's hardware. [6] [7]

The UR route uses WSG50 configuration and a camera capture chain. The Franka instructions describe a separate Polymetis server on a real-time-kernel computer and different load settings. Neither route was installed or connected to hardware by this catalog. [1] [5] [8]

The cup-policy download command names cup_wild_vit_l_1img.ckpt, while the shown evaluation command uses cup_wild_vit_l.ckpt. Resolve that filename/configuration mismatch before assuming the example runs unchanged. The README also reports poor performance in direct sunlight for that policy; the project page's broader lighting-robustness claims should not erase this specific limitation. [1] [6]

The repository is MIT-licensed. That license is not automatically assigned to external CAD, hardware-guide media, datasets, checkpoints or the SLAM fork. This profile establishes source-researched workflow coverage, with no fabrication, training, flashing or physical performance test claimed. [1] [3] [5] [9]

What we checked

Source research only; no installation, physical build or performance test performed. No affiliate links added.

Not yet established

  • Requirements: tools
  • A step-by-step setup summary

Updates

Upstream last updated
May 26, 2026 [2]
Latest release
Not established
Profile updated
Oct 25, 2026
Sources last checked
Oct 1, 2026
Next review
Oct 31, 2026
Reviewed revision
d095ba9590df789df5189eea5ee7e431689038a6 (commit snapshot; no published stable release found) · combined hardware and software repository

Source files and licenses

Available project sources support research coverage; complete build and commercial qualification remain separate.

LayerAvailability / license
mechanical designavailable · License not fully assessed
Official hardware guide links gripper CAD, separate soft-finger CAD and deployment mounts. External CAD payloads, exact revisions and hardware reuse licenses were not independently preserved or established. [5]
electronicsnot found · License not fully assessed
Off-the-shelf camera and robot electronics are described; no complete custom electronics fabrication package established in the selected sources. [5] [9]
firmwarenot found · License not fully assessed
GoPro Labs, robot and gripper firmware are external dependencies; complete matched firmware sources and licenses were not collected. [1] [5]
softwareavailable · MIT
Selected repository code; external SLAM fork, robot packages and dependencies retain separate terms. [1] [3]
model weightsnot found · License not fully assessed
Official README links external pretrained checkpoints and datasets, but their payloads and licenses were not independently preserved in this review. This is not a claim that the artifacts are unavailable upstream. [1] [9]
datasetsnot found · License not fully assessed
Official README links external pretrained checkpoints and datasets, but their payloads and licenses were not independently preserved in this review. This is not a claim that the artifacts are unavailable upstream. [1] [9]
documentationavailable · MIT
Repository documentation only; external hardware guide, CAD, media and data are not assigned the software license. [1] [3]

References

Each source links to the exact revision we read. “Used for” shows where this page relies on it.

Project sources

Videos

  • S010 UMI - 3D Printing Tutorial (YouTube video)
    Video from the project’s own channel (Robotics & Embodied Artificial Intelligence Lab), linked from the project’s own documentation. Not reviewed by us for accuracy.
    Revision, date & where cited

    YouTube video Robotics & Embodied Artificial Intelligence Lab · youtube.com · Accessed Oct 4, 2026

  • S011 UMI - Assembly Tutorial (YouTube video)
    Video from the project’s own channel (Robotics & Embodied Artificial Intelligence Lab), linked from the project’s own documentation. Not reviewed by us for accuracy.
    Revision, date & where cited

    YouTube video Robotics & Embodied Artificial Intelligence Lab · youtube.com · Accessed Oct 4, 2026

Correction guidance · Discussion is not enabled · Profile format 0.6.0-draft.1

Change history

2026-10-04 — Added 2 YouTube reference(s) (S010, S011): official channel or videos linked from the pinned upstream sources. Existing claims, source-check dates and evidence are unchanged.

2026-10-25 — Added an attributed upstream image with linked reuse terms; technical evidence dates unchanged.