Grabette Opens Robot Data Collection to Anyone

Robot learning is running into a data bottleneck. Modern systems already have capable policy architectures—including transformer-based vision-language-action models, diffusion and flow-matching policies, and world models—as well as sufficient computing power for training. The harder problem is obtaining enough varied, real-world demonstrations of physical manipulation. Conventional collection usually requires an actual robot and a teleoperation setup, making the process costly, tiring, technically demanding, and difficult to reproduce across the many tasks and environments needed for broadly useful robot policies.

Grabette approaches the problem by separating demonstration recording from robotic execution. Instead of remotely controlling a robot, a person performs a task directly with a handheld gripper equipped with cameras and sensors. The system captures what the hand and gripper do, reconstructs the device’s six-degree-of-freedom pose—the three-dimensional position and orientation of the tool—and combines that trajectory with visual observations and the gripper state. The result is intended to be a clean dataset that can later be used to train a robot. The central idea is that useful robot-manipulation demonstrations can be collected without having a robot present during recording.

The project is presented as an open, comparatively low-cost system that people can assemble from purchasable components and operate outside a specialized laboratory. Its larger ambition is to make recording a manipulation demonstration closer to shooting an ordinary video. If the process becomes accessible to many contributors, the creators believe Grabette could help seed a large, open, collaborative manipulation dataset containing more tasks and environments than a single laboratory could realistically collect on its own.

Grabette draws direct inspiration from Stanford’s Universal Manipulation Interface, or UMI. That project demonstrated a similar recipe: use a handheld gripper with a fisheye camera to record real-world demonstrations, recover camera trajectories through simultaneous localization and mapping, and train visuomotor policies from the resulting data. The authors also identify commercial or closed-source alternatives, including Agibot’s MEgo gripper, Genrobot’s DAS gripper, and Sunday Robotics’ skill-capture glove. Grabette’s distinction is its emphasis on openness, accessible construction, standard components, and integration with an existing open machine-learning ecosystem.

The handheld device carries two cameras with deliberately separate responsibilities. A low-cost, wide-angle fisheye camera records a context-rich view resembling the wrist-mounted camera that a robot policy would use. An OAK-D RGBD camera, which captures color and depth, supports more robust six-degree-of-freedom tracking. The device also contains an inertial measurement unit and magnetic encoders that measure the gripper’s joint state. Dividing visual observation and motion tracking between dedicated sensors is meant to give the learned policy an appropriate viewpoint without forcing the same inexpensive camera to perform the more demanding tracking task.

Grabette records demonstrations, but it does not execute the learned movements. That role belongs to Gripette, a related motorized end effector designed to be mounted on a real or simulated robot arm. The two devices share the same underlying hardware approach: Grabette is the handheld capture instrument with cameras, an IMU, and a gripper, while Gripette combines a camera with two servomotors to reproduce learned actions. The stated bill-of-materials cost is approximately €490 for Grabette and €120 for Gripette. A matching Gripette is still required on the target arm even though the recorded data and capture process are otherwise designed to remain independent of a particular robot platform.

The release covers more than the physical recorder. It includes CAD and production files for both devices, Raspberry Pi software for the on-device capture service, a post-processing pipeline that can run locally or through a Hugging Face Space, and an example downstream workflow using standard LeRobot training with OpenArm evaluation. The hardware relies on ordinary components rather than a proprietary pipeline: a Raspberry Pi, a standard Raspberry Pi camera, an off-the-shelf OAK-D depth camera, and magnetic encoders. The intention is to let users reproduce, inspect, modify, and extend the system without being tied to a vendor-specific software fork.

The recorded representation is robot-agnostic at the dataset level. Each demonstration stores a camera-local Cartesian pose with six degrees of freedom together with the gripper state. Processed episodes are exported as standard LeRobot datasets and can be shared through the Hugging Face Hub. In principle, that format allows the same demonstrations to support different robot arms and learning methods, although physical execution still depends on equipping the arm with the corresponding Gripette end effector.

The capture workflow is designed for users without specialist experience. Pressing a button begins synchronized recording from the observation camera, the tracking camera’s color, depth, and IMU streams, and the gripper encoders. All signals use a shared clock so that images, motion estimates, and joint values remain aligned. Pressing the button again ends the episode and saves its raw data locally on the Raspberry Pi. The user then opens a browser-based dashboard, selects the episodes to include, and starts post-processing with one click.

During processing, the grabette-slam Hugging Face Space uses the RTAB-MAP library to perform simultaneous localization and mapping and reconstruct the device trajectory. It also checks the result for discontinuities or tracking loss, helping prevent corrupted motion paths from entering the final dataset. Accepted episodes are converted into LeRobot format and uploaded to the user’s Hub space, where each episode can be inspected with the LeRobot visualizer. By connecting synchronized capture, trajectory validation, standardized conversion, browser-based processing, and dataset sharing, Grabette aims to shorten the path from demonstrating a task to obtaining training-ready robot data.

The creators say they have developed the system for several months and now consider it usable enough for a public release. The project does not eliminate the need for robot hardware at the training-evaluation or deployment stage, and its broad impact will depend on how reliably demonstrations transfer across robots and environments. Nevertheless, it targets one of embodied AI’s most practical constraints: collecting diverse physical experience at a scale that traditional laboratory teleoperation workflows struggle to reach.