UMI gripper profile: collecting robot data without a robot

Contents

1. How it works
2. Published evidence
3. Strengths and limits
4. Pricing
5. Questions to settle before building
6. Sources

Type	Handheld capture gripper
Founded	Published at RSS 2024
Pricing	Not published
Website	umi-gripper.github.io

Some teams looking for teleoperation hardware need demonstrations, not control. For them the Universal Manipulation Interface, a finalist for the best systems paper award at RSS 2024, removes the robot from data collection altogether⁠1Universal Manipulation Interface: in-the-wild robot teaching without in-the-wild robotsChi et al. (RSS 2024).

Best for: Collecting many gripper demonstrations quickly, in real places, with no robot on site.

Not the right fit for: Live control of a robot: UMI records demonstrations, it does not drive anything.

How it works

The hardware is a handheld parallel-jaw gripper with a GoPro camera mounted on it. A person carries it through the task. The recording captures the visual context, the gripper’s motion and fine changes in gripper width, and the project’s software checks whether each demonstration is feasible for a given robot. The project page links a codebase, a hardware guide and printing and assembly tutorials⁠1Universal Manipulation Interface: in-the-wild robot teaching without in-the-wild robotsChi et al. (RSS 2024).

Published evidence

The authors report about 30 seconds per demonstration. On a cup arrangement task they measured 111 demonstrations an hour with the UMI gripper, 231 with a bare human hand and 35 with space-mouse teleoperation. The same policy ran on UR5e and Franka arms, and the authors state it can run on any robot with a parallel jaw stroke over 85 mm⁠1Universal Manipulation Interface: in-the-wild robot teaching without in-the-wild robotsChi et al. (RSS 2024). These are the authors’ figures.

Strengths and limits

Strengths

• About three times the throughput of space-mouse teleoperation on the authors’ cup task.
• Hardware-agnostic: the authors state the policy can run on any robot with a parallel jaw stroke over 85 mm.
• A demonstration takes about 30 seconds, per the authors, and needs no robot in the room.

Limits

• Not teleoperation. It cannot take over a deployed robot or correct one mid-task.
• Limited to parallel-jaw gripping.
• No price is stated on the project page.

Choose another option if: You need a person to control the robot in real time, for interventions or for data on the robot’s own body (any other option here).

Pricing

Not published on the project page.

Questions to settle before building

• Does the task need live control of the robot at any point, such as taking over a deployed unit?
• Does our gripper meet the stroke requirement?
• Who will carry the gripper, where, and on what protocol?
• How will we check that the collected demonstrations run on our arm?

See how UMI compares with the other options in the rig comparison table.

Sources

1. Universal Manipulation Interface: in-the-wild robot teaching without in-the-wild robots. Chi et al. (RSS 2024). Accessed 11 Nov 2026.

Further reading

• Buyer's guide · Updated quarterlyTeleoperation hardware in 2027: controllers, kits and rigs comparedNine ways to put a robot under a person's control, from a $230 pair of printed arms to a $24,000 bimanual station, compared on what they cost and what they record, followed by the line most hardware budgets leave out, the operators.
• Guide · Leader arms and kitsRobot teleoperation kits: SO-101, GELLO and the ALOHA line comparedA leader arm is the simplest teleoperation controller and the one that records the cleanest data. The kits differ by a factor of a hundred in price. This guide sets out what the money buys.
• ProfileAirExo profile: a low-cost exoskeleton for whole-arm teleoperationA wearable leader arm at about $300 per side that can also record demonstrations with no robot in the room.