Grabette: an open system to record robot-manipulation data

The primary obstacle hindering the advancement of robotics today is not the sophistication of our neural architectures or the availability of high-end compute. Whether we are discussing transformer-based Vision-Language-Action (VLA) models, diffusion policies, or cutting-edge world models, the industry has reached a point where the bottleneck is fundamentally rooted in data. Specifically, we suffer from a profound scarcity of diverse, high-quality, real-world manipulation datasets.

Historically, collecting this data has been a gatekept, expensive, and logistically exhausting endeavor. It typically requires access to specialized robotic hardware, complex teleoperation rigs, and hours of tedious manual labor. Scaling this process to cover the vast array of tasks and environments necessary for general-purpose robotics has proven nearly impossible for individual labs.

Today, that paradigm shifts. Introducing Grabette, an open-source, low-cost system designed to democratize the collection of manipulation data. By removing the need for a robot during the recording phase, Grabette allows anyone to capture complex tasks using nothing more than a handheld gripper and a camera, turning human demonstrations into robot-ready datasets with unprecedented ease.

Standing on the Shoulders of UMI

The philosophy behind Grabette is deeply rooted in the Universal Manipulation Interface (UMI), the groundbreaking research from Stanford University. UMI demonstrated that a handheld gripper equipped with a fisheye camera could record demonstrations "in the wild," utilizing SLAM (Simultaneous Localization and Mapping) to recover camera trajectories and train visuomotor policies.

While other proprietary solutions—such as Agibot’s MEgo, Genrobot’s DAS, and Sunday Robotics’ skill capture glove—have emerged, they often remain locked behind closed ecosystems. The team behind Grabette, led by Pollen Robotics, sought to lower the barrier to entry significantly. Their goal was to bridge the gap between "I have a task" and "I have a trained model" by integrating seamlessly into the modern open-source robotics stack, including LeRobot and the Hugging Face Hub.

"The bottleneck isn’t the model. It’s the data. We want Grabette to seed a large, open, collaborative manipulation dataset. One no single lab could ever build alone."

Meet the Hardware: Grabette and Gripette

Grabette is the result of months of iterative development, designed to be accessible, modular, and highly effective. The system consists of two primary components that share the same hardware DNA, ensuring that what a human records can be faithfully executed by a machine.

The Grabette Handheld Device

This is the primary data-collection tool. It is a handheld gripper instrumented with two distinct cameras, each serving a specific purpose:

• Wide-Angle Fisheye Camera: Provides the context-rich, wrist-camera-style perspective essential for training robust policies.
• RGBD Camera: Handles the heavy lifting of precise, robust 6-DoF (degrees of freedom) tracking.

With a Bill of Materials (BOM) cost of approximately 490€, the device is accessible to researchers, hobbyists, and students alike.

The Gripette Robotic Twin

To close the loop, the team developed Gripette, a motorized gripper designed to act as the robotic end-effector twin. With a BOM cost of roughly 120€, it features a camera and two servomotors, allowing it to execute the movements learned during the training phase on a real or simulated robot arm.

A Two-Step Pipeline: From Hand to Dataset

The brilliance of Grabette lies in its streamlined workflow. The system is designed to take a user from a physical demonstration to a training-ready dataset in just two steps, requiring no specialized expertise.

1. Record: By pressing a single button, the system synchronizes data from the observation camera, the tracking camera (color, depth, and IMU), and the gripper’s encoder joint values. This shared clock ensures that the resulting data is perfectly aligned. Once the episode concludes, the data is saved locally on the integrated Raspberry Pi. 2. Process: Users simply open the Grabette dashboard in their web browser. With a single click, the system triggers a post-processing pipeline that uploads the episodes to the Hugging Face Hub. The grabette-slam space then utilizes the RTAB-MAP library to verify trajectory integrity, converts the data into the standard LeRobot format, and prepares it for immediate training.

Robot-Agnostic and Community-Driven

A core tenet of the Grabette project is its commitment to being robot-agnostic. The data format does not assume the use of any specific robotic arm; instead, it stores demonstrations as camera-local 6-DoF Cartesian poses combined with gripper states. This ensures that the same data can be utilized across different hardware platforms and learning methodologies.

To prove the efficacy of the system, the team provided a reference implementation. By recording 200 demonstrations, they trained a Diffusion Policy (using a ResNet18 encoder and a DDIM scheduler) that fits comfortably on a single consumer-grade GPU. This policy was then successfully evaluated on an OpenArm 7-DoF robot, demonstrating that the system is not just a concept, but a functional end-to-end pipeline.

Join the Movement

The data bottleneck in robotics will not be solved by a single institution; it requires a global community of contributors recording tasks in diverse environments. By building a Grabette, recording tasks, and sharing them on the Hugging Face Hub, contributors are helping to dismantle the barriers that have historically kept robot learning locked behind expensive, proprietary hardware.

Key Features of the Grabette Ecosystem:

• Fully Open Source: Access to all CAD files, production blueprints, and software repositories.
• No Vendor Lock-in: Built using off-the-shelf components like Raspberry Pi, standard Pi cameras, and OAK-D depth sensors.
• Browser-Based Processing: No complex local installations required; the processing pipeline is accessible directly via the web.
• Standardized Output: Compatible with the LeRobot stack, making it easy to integrate into existing research workflows.

As the project looks toward the future, the team is already working on Casquette, a head-mounted POV device designed to complement Grabette by capturing egocentric data. However, the most critical next step lies with the community. The tools are ready, the pipeline is open, and the dataset is waiting to be built. It is time to start recording.