Grabette is an open-source, low-cost handheld gripper and processing ecosystem that lets anyone record real-world manipulation demonstrations in minutes and automatically convert them into robot-ready datasets.
Why this matters
The bottleneck in robot learning today is data, not models. We have powerful policy architectures (transformer-based VLAs, diffusion and flow-matching policies, and world models) and GPUs to train them, but large, diverse real-world manipulation datasets are scarce. Teleoperating robots to collect data is often expensive and hard to scale, requiring hardware, labs or teleoperation rigs.
Grabette’s premise is simple: you do not need a robot to collect robot data. A human hand, a gripper, a camera, and a method to recover the hand’s 6‑DoF trajectory are enough. Capture the demonstration and you have data a robot can learn from.
Roots in UMI
Grabette is directly inspired by the Universal Manipulation Interface (UMI) from Stanford: a handheld gripper with a fisheye camera that records in-the-wild demonstrations, recovers camera trajectories with SLAM, and is used to train visuomotor policies. UMI showed the concept works; other closed-source devices exist as well (for example Agibot’s MEgo, Genrobot’s DAS, and Sunday Robotics’ capture glove).
Grabette aims to make the process effortless and reduce the barrier from “I have a task” to “I have a trained model.”
System overview
Grabette is built into a modern open ecosystem: LeRobot for dataset format, the Hugging Face Hub for sharing, and a processing pipeline you can run from your browser with nothing to install. The device can be built on a workbench, used in the field, and used to contribute data.
Hardware and components
- Grabette: the handheld demonstration device (camera + IMU + gripper). BOM cost ~490 €.
- Gripette: the motorized gripper, the robotic twin of Grabette (camera + two servomotors). BOM cost ~120 €.
Grabette carries two cameras with distinct roles: a cheap wide fisheye for the context-rich, wrist-camera-like view the policy benefits from, and an RGBD camera (e.g., OAK-D) that performs robust 6‑DoF tracking. Other parts are standard and off-the-shelf: a Raspberry Pi, IMU, magnetic encoders — components anyone can order.
Everything is open source
The project publishes CAD and production files for both Grabette and Gripette, the on-device Raspberry Pi capture software, the processing pipeline (runnable locally or via a Hugging Face Space), and an example downstream stack (stock LeRobot training + OpenArm evaluation).
Robot-agnostic data format
The capture format does not assume a particular robot arm: demonstrations are stored as camera-local 6‑DoF Cartesian poses plus gripper state, and the output is a standard LeRobot dataset on the Hugging Face Hub. This allows the same data to drive different robots and learning methods; executing learned behaviors does require a matching Gripette gripper on the target arm.
From your hand to a dataset in two steps
The release makes it possible for anyone to quickly go from recording a demonstration to getting a training-ready dataset.
- Record
- Press the button to start recording. Observation camera, tracking camera (color, depth, and IMU), and gripper encoder joint values are recorded simultaneously with a single shared clock for synchronization. Press the button again to stop the episode; data is saved locally on the Raspberry Pi.
- Process from your browser
- Open the Grabette dashboard, select episodes to add to a dataset, and click to start post-processing.
- Episodes are uploaded to the Hugging Face Hub.
- The grabette-slam Space runs SLAM using the RTAB-MAP library and verifies the trajectory (checking for jumps or tracking loss).
- Episodes are converted to LeRobot format.
- A new dataset is uploaded to your space, and you can inspect episodes with the LeRobot visualizer.
After these steps the data is ready for training.
An example end-to-end demonstration
To show the full loop, the project ships a complete example. Using around 200 recorded demonstrations (the dataset is available as an HF Dataset), the LeRobot stack trains a Diffusion Policy (ResNet18 + SpatialSoftmax encoder, DDIM scheduler, 6‑D rotation actions) that fits on a single consumer GPU. The policy is evaluated on an OpenArm 7‑DoF arm with the Gripette gripper, driven via a gRPC API. The example policy and dataset are available on the Hugging Face Hub.
How you can help
The data bottleneck will be solved by a community-wide effort, not a single lab. Build a Grabette, record tasks you care about, and share them on the Hub. Each episode increases the size and diversity of the open dataset and lowers the barrier to robot learning.
What’s next
This release is a starting point. The Grabette project will continue to evolve: the team is already working on additions such as Casquette, a head-mounted POV device to complement Grabette for egocentric capture (still a work in progress). The most important next step is community participation: start recording and help build the dataset.
Credits and resources
The project is developed by Pollen Robotics. The GitHub repository, Hugging Face Spaces, and example stacks are published so anyone can build, process, and contribute data.
