Hugging Face has released Grabette, an open-source system for recording robot manipulation data that aims to solve one of the field's most persistent bottlenecks: the shortage of large, diverse, real-world training data.
The device is deliberately simple. It consists of a handheld gripper with two cameras: a wide fisheye for context-rich wrist-camera-style views, and an RGBD camera for robust 6-DoF tracking. A Raspberry Pi handles recording and local storage. The total bill of materials is approximately 120 euros for the motorized gripper variant, or less for the passive recording version.
Users press a button to start recording. The system captures synchronized data from the observation camera, the tracking camera (color, depth, and IMU), and the gripper's encoder values. Press again to stop, and the episode saves locally. Post-processing happens through a browser dashboard: select episodes, run SLAM verification to check for tracking jumps, and upload the resulting dataset to the Hugging Face Hub in standard LeRobot format.
The Data Problem
Robot learning has capable policy architectures and enough GPU compute to train them. What it lacks is data. Teleoperating robots to collect demonstrations is expensive, tedious, and difficult to scale across the variety of tasks and environments needed for generalization. Grabette's bet is that if recording a demonstration is as easy as shooting a video, anyone can contribute to a shared dataset that no single lab could build alone.
The design is directly inspired by Stanford's UMI gripper, which proved that handheld devices with SLAM-recovered trajectories can train visuomotor policies. Grabette extends this with modern open tooling: LeRobot for datasets, the Hugging Face Hub for sharing, and a processing pipeline that requires no installation.
Robot Counterpart: Gripette
To close the loop from data to deployment, Hugging Face is also releasing Gripette, a motorized gripper that mounts on robot arms and mirrors Grabette's form factor. The same LeRobot dataset format drives both training and execution, meaning data collected with Grabette can train policies that run on any arm equipped with Gripette.
Hugging Face shipped an end-to-end example using a Diffusion Policy with ResNet18 encoder, trained on a single consumer GPU, and deployed on an OpenArm 7-DoF arm with Gripette over a gRPC API.
Open Dataset Ambition
The broader goal is a community-built open manipulation dataset. Hugging Face is encouraging anyone with a Grabette to record tasks they care about and share them on the Hub. The company says future updates will include Headette, a head-mounted POV device for egocentric capture.
The release reflects a strategic bet that open data, not just open models, is what will democratize robot learning. If the approach works, the barrier to entry for training manipulation policies drops from "access to a robotics lab" to "access to a 3D printer and some off-the-shelf parts."