Let’s talk
← Back to home

UMI / DATA ACQUISITION

UMI data collection
for manipulation.

Human demonstrations captured with handheld grippers. We collect UMI datasets that pair what the gripper sees with how it moves, scoped around the manipulation tasks your policy needs to learn.

Build your dataset How a collection runs

01 / WHAT IT IS

Human demonstrations.
Physical action.

UMI, the Universal Manipulation Interface, is a data collection approach in which a person performs a manipulation task using a handheld gripper fitted with a camera and pose tracking. The demonstrator’s own hand does the work; the device records what a robot gripper would have seen and how it would have moved.

Because no robot is needed during collection, UMI demonstrations can be gathered quickly, in many locations, and with natural human dexterity. The resulting data maps closely onto what a robot arm with a similar gripper needs as input and output, which is why it is widely used to train manipulation policies from demonstration.

Our UMI collections are aligned to your target setup. Gripper geometry, camera placement, and the tasks demonstrated are agreed in the brief so that the data transfers to the robot you are actually training.

02 / WHAT WE CAPTURE

What a UMI
collection contains.

The device configuration follows your target robot. A typical UMI collection includes:

Gripper-mounted video
Wrist-view footage from the camera on the handheld gripper, capturing the object and the gripper fingers throughout the demonstration.
Gripper pose
The six-degree-of-freedom trajectory of the gripper through the task, from the tracking method agreed in the brief.
Gripper width
Opening and closing of the fingers over time, aligned with the video and pose streams.
Episode segmentation
Each demonstration cut to your task definition, with start and end states noted.
Variation and quality tags
Object, layout, and site metadata per episode, plus review flags for demonstrations that fell outside the brief.

03 / WHAT IT IS FOR

Where UMI data
earns its place.

Training manipulation policies from demonstration, where the policy input is the wrist camera view and the output is gripper motion and width.

Covering object and scene variation cheaply: because collection does not need a robot, a UMI dataset can span more objects, surfaces, and sites than a lab teleoperation session.

Complementing teleoperation data on the same task, giving a policy both human-dexterous demonstrations and embodiment-specific ones.

04 / WHERE WE COLLECT

Collected across the
Middle East and Central Asia.

Telepath is based in Dubai and runs collections in real workplaces across the United Arab Emirates and the CIS, with sites in Kazakhstan, Uzbekistan and Tajikistan. The data is built for robotics and physical-AI teams in the United States, led by the San Francisco Bay Area.

Emirates
Dubai, Sharjah and Ras Al Khaimah.
Site types
Laundries, perfume and cosmetics plants, cable factories, furniture factories, plus logistics, hospitality and retail sites.
Languages
Briefs and on-site instructions in English, Arabic and Russian.
Why it matters
Real objects, real layouts and real interruptions, so the variation your model meets in deployment is already in the training data.
Scope a collection Offer a site or record at yours

05 / EXAMPLE TASKS

Briefs we can scope
with this method.

These example briefs suit handheld-gripper capture, particularly tasks with clear grasps, placements, and orientations. Custom tasks are welcome.

EXAMPLE COLLECTION BRIEFS · CUSTOM TASKS WELCOME

06 / HOW A COLLECTION RUNS

Four stages,
one shared brief.

A UMI collection follows our four-stage process. The Define stage fixes the gripper configuration and demonstration protocol so the data matches your robot.

  1. DefineStart with the capability.

    Agree the task, the gripper geometry and camera placement to match your target robot, the demonstration protocol, and the variations to cover.

    YOUR OUTPUT / A SHARED COLLECTION BRIEF

  2. CollectBring the brief into the world.

    Run demonstrations at the agreed sites with trained demonstrators, checking tracking quality and episode completeness on site.

    YOUR OUTPUT / TASK-ALIGNED DEMONSTRATIONS

  3. ReviewMake every delivery intentional.

    Segment and tag episodes, verify pose and video alignment, and review each demonstration against the brief. Re-collect where tracking or coverage falls short.

    YOUR OUTPUT / REVIEWED COLLECTION DATA

  4. DeliverFit your training workflow.

    Deliver synchronised video, pose, and gripper-width streams per episode, with manifest and metadata in the agreed structure.

    YOUR OUTPUT / DATA STRUCTURED FOR YOUR TEAM

FAQ

UMI data, answered.

What is UMI in robotics data collection?

UMI stands for Universal Manipulation Interface. It is a way to collect manipulation demonstrations using a handheld gripper with a camera and pose tracking, so a person can demonstrate a task and produce data that looks like what a robot gripper would record.

Do we need our own robot to collect UMI data?

No. UMI data collection does not require a robot, because the demonstrations are captured with a handheld gripper device rather than a robot arm. The device is configured to match your target gripper geometry and camera placement, so the demonstrations transfer to your platform when you train on them.

How does UMI data differ from teleoperation data?

UMI data comes from a human hand moving a handheld gripper. Teleoperation data comes from a real robot being driven by an operator. UMI is faster to collect and covers more variation; teleoperation captures the exact embodiment. Many teams use both.

Which tasks work well with UMI?

UMI data collection works well for manipulation tasks with clear grasps, placements, and orientation changes: parts kitting, product facing, product sorting, tool handling, and similar single-gripper work. Long bimanual sequences can be scoped with two devices where the brief calls for it. Tasks that depend on the exact robot embodiment are usually better served by teleoperation data.

What do you deliver?

A UMI data collection is delivered as per-episode video, gripper pose, and gripper width, synchronised and segmented to your task definition. Each delivery includes a manifest, variation metadata for every episode, and a coverage report against the brief. File formats and dataset layout follow what was agreed at the Define stage.

FROM THE WORLD. INTO YOUR MODEL.

Put a gripper
in human hands.

Build your dataset