You don't need a robot to train one: open-source handheld gripper aims to crowdsource robot learning data

You Don't Need a Robot to Train One

Pollen Robotics, now part of Hugging Face, has a claim that cuts against the direction of the robotics industry: the bottleneck in robot learning is not AI model capability. It is data. And the fix is not a billion-dollar lab. It is a €490 handheld gripper you can build on a workbench.

The device is called Grabette. You hold it like a tool, perform a manipulation task with your own hand, and press a button. Two onboard cameras, an inertial measurement unit, and a magnetic encoder on the gripper jaw record everything. Post-processing software, which runs in a browser with nothing to install, reconstructs the trajectory of your hand movement using SLAM and packages the result as a standard LeRobot dataset on the Hugging Face Hub 1.

No robot arm required. No lab. No teleoperation rig. The whole point is to collapse the gap between wanting to demonstrate a task and having data a robot can learn from 1.

That matters because of what Grabette is designed to seed. The team's explicit goal is a large, open, collaborative manipulation dataset that no single lab could build alone 1. Think of what happened with maps. OpenStreetMap started with volunteer GPS traces and grew into the geographic backbone that apps and automakers build on, eroding the moats of proprietary map providers. Grabette is a bid to do the same for robot data. If a globally distributed community of hobbyists, researchers, and tinkerers each contributes a few hundred demonstrations, the resulting dataset could exceed what any well-capitalized robotics company collects behind closed doors.

That is the signal for anyone building or backing robotics companies. In language models, open datasets like Common Crawl compressed the data advantage of any single lab. They did not prevent proprietary models from existing, but they shrank the gap between those who could pay for data and those who could not. The same dynamic could play out in robotics.

Grabette is directly inspired by the Universal Manipulation Interface, or UMI, a handheld gripper developed at Stanford that proved demonstrations recorded in everyday environments could train robot policies 1. UMI was a research proof of concept. Grabette is built for adoption. The bill of materials for the handheld recorder is roughly €490, and the motorized counterpart that attaches to a robot arm, called Gripette, costs about €120 in parts 1. Both designs are fully open source under an Apache-2.0 license, with CAD files, firmware, and processing code on GitHub 2.

Closed-source competitors already exist. The team names Agibot's MEgo gripper, Genrobot's DAS gripper, and Sunday Robotics' skill capture glove as similar devices, all proprietary 1. Grabette's pitch is accessibility: standard sensors you can order online, a browser-based processing pipeline, and integration with LeRobot so data flows directly into existing training workflows 1.

The hardware is deliberately ordinary. A Raspberry Pi, a standard Pi camera, an off-the-shelf OAK-D depth camera, magnetic encoders. The design uses two cameras for distinct roles: a cheap fisheye gives the learning policy a wrist-camera-style view of the scene, while the RGBD camera handles the precision work of tracking position and orientation in six degrees of freedom 1. The data format is robot-agnostic: demonstrations are stored as camera-local poses plus gripper state, so the same recordings can train different robots 1.

To prove the loop works end to end, the team trained a Diffusion Policy on 200 recorded demonstrations and evaluated it on an OpenArm 7-DoF arm equipped with the Gripette gripper 1. The model fits on a single consumer GPU. But the team is clear that the release is the recording system, not the trained policy. The point is that anyone's Grabette data will work with any method that consumes LeRobot datasets 1.

The company now sits inside Hugging Face, and Grabette reflects that position: infrastructure that lets anyone contribute to robot learning, not just buy a robot. Casquette, a head-mounted capture device still in development, would extend the system to first-person demonstrations 1.

The underlying thesis is worth quoting directly: "The bottleneck isn't the model. It's the data" 1. If that is correct, and if enough people build a Grabette and start recording, the question for every robotics company building a proprietary data advantage becomes uncomfortable. How long does that advantage last when anyone with €490 in parts and a workbench can contribute to a shared dataset anyone can use?