Grabette: An Open-Source System for Recording Robotic Manipulation Data
And building a shared dataset together.
- The bottleneck isn’t the model, it’s the data.
- Standing on the shoulders of UMI
- Meet Grabette
- Built for everyone
- From your hand to a dataset in two steps
- What can you do with the data? Here’s one example.
- Now it’s your turn
- What’s next
Simply hold the gripper, record your own manipulation tasks in minutes, automatically convert them into robot-ready datasets, and help grow an open, collaborative dataset for robot learning.
The bottleneck isn’t the model, it’s the data.
Robot learning has a supply problem. We already have powerful enough policy architectures (Transformer-based VLAs, diffusion and Flow Matching policies, even world models) and plenty of GPUs to train them. What we lack is large-scale, diverse, real-world manipulation data.
Using teleoperated robots to collect data is both expensive and tedious: first, it requires a robot. And depending on the teleoperation method, if data collection takes hours, it becomes cumbersome for the user and introduces significant hardware and logistical challenges. Scaling this up to match the wide variety of tasks and environments needed is incredibly difficult.
But you don’t need a robot to collect robot data. All you need is a human hand, a gripper, a camera, and a way to recover the 6-DoF trajectory of the hand’s motion. Capture the demonstration, and you have data the robot can learn from.
That is what we are releasing today: Grabette, an open-source, low-cost system for recording manipulation data. Pick it up, record tasks with your own hand, and get a clean, robot-ready dataset. No robot, no lab, no teleoperation rig required.
The bigger goal lies here: if recording a demonstration is as simple as shooting a video, then anyone can contribute data. We want Grabette to foster a large, open, collaborative manipulation dataset—something no single lab could ever build alone.
Standing on the shoulders of UMI
Grabette is directly inspired by Stanford’s Universal Manipulation Interface (UMI): a handheld gripper equipped with a fisheye camera that records demonstrations in the wild, recovers camera trajectories via SLAM, and trains visuo-motor policies accordingly.
UMI proved that this methodology works. Other closed-source devices exist, such as Agibot’s MEgo gripper, Genrobot’s DAS gripper, and Sunday Robotics’ skill capture glove.
Our goal is to make it as easy to use as possible, lowering the barrier from “I have a task” to “I have a trained model” to the absolute minimum.
Grabette is built on top of a modern open ecosystem: LeRobot for datasets, Hugging Face Hub for sharing, and a processing pipeline that runs directly in your browser without installing anything. Grabette is a tool that anyone can build on their workbench, use in the field, and contribute data to.
Meet Grabette
We have been developing Grabette for a few months now and feel it is mature enough to share. We are excited to release it today!
Grabette is a handheld gripper equipped with everything needed to reconstruct manipulation demonstrations.
It features two cameras, each with a distinct role. Splitting these roles is intentional: a cheap, wide-angle fisheye camera provides the policy with the context-rich perspective it needs (similar to a wrist camera), while an RGBD camera handles the heavy lifting of robust 6-DoF tracking.
While Grabette records the data generated by a user performing a task, it relies on a corresponding robot-side end-effector to execute the learned actions after training. Enter its robotic twin: Gripette, the robotic arm end-effector version of Grabette.
This family of devices shares the same hardware DNA:
-
Grabette: Handheld demonstration capture device (cameras + IMU + gripper, BOM cost ~€490)
-
Gripette: Motorized gripper (camera + two servos, BOM cost ~€120) used for closed-loop control on real or simulated robotic arms
Grabette handheld recording device (left) and Gripette robotic gripper (right)
Built for everyone
Everything is open-source
- Hardware: CAD and manufacturing files for Grabette and Gripette
- Capture Service: On-device Raspberry Pi software
- Processing Pipeline: Can be run locally or online via our Hugging Face Space
- Downstream Example Stack: Standard LeRobot training pipeline + OpenArm evaluation, provided as a reference
Components. Everything is built using standard, off-the-shelf sensors—no closed pipelines, no vendor lock-in. A Raspberry Pi, a standard Pi camera, an off-the-shelf OAK-D depth camera, and magnetic encoders. The core goal is to allow anyone to assemble it using parts they can order directly online.
Robot-agnostic by design. The capture pipeline and data format do not assume a specific robotic arm. Demonstrations are stored as 6-DoF Cartesian poses in the camera’s local frame, along with the gripper state. The output is a standard LeRobot dataset on the Hugging Face Hub, meaning the same data can drive different robots and learning methods. However, you will still need to use the corresponding Gripette gripper on your robotic arm.
From your hand to a dataset in two steps
This release makes it fast for anyone to go from “I want to demonstrate a task” to “I have a dataset ready for training,” without requiring any prior experience.
1. Record
Once the button is pressed, data from the observation camera, tracking camera (color, depth, and IMU), and gripper encoder joint values are recorded simultaneously, using a shared clock to ensure proper synchronization. Pressing the button again stops the episode, and the data is saved locally on the Raspberry Pi.