Use a handheld gripper to record unique manipulation tasks in minutes and automatically transform them into robot-ready datasets, helping to grow open and collaborative datasets for robot learning.
The bottleneck is not the model. It’s data.
Robotic learning has a supply problem. We have capable policy architectures (transformer-based VLAs, diffusion and flow matching policies, and even world models) and GPUs to train them. What we lack is large and diverse real-world operational data.
Remotely controlling robots for collection is expensive and demanding. First, we need a robot. Additionally, some remote control methods can take hours of data collection and involve significant hardware and logistical challenges, which can be tedious for users. This is difficult to scale due to the wide variety of tasks and environments required.
However, you don’t need a robot to collect robot data. Just a human hand, a gripper, a camera, and a way to recover the 6-DoF trajectory of what the hand did. Capturing demonstrations provides data for the robot to learn from.
That’s what we’re releasing today. Grabette is an open, low-cost system for recording operational data. Pick it up, record tasks with your own hands, and get back a clean, robot-ready dataset. There are no robots, labs, or teleoperation equipment.
And that’s the bigger goal. If recording a demo is as easy as shooting a video, then anyone can contribute. We want Grabette to seed large, open collaborative datasets. No laboratory can ever be built alone.
stand on the shoulders of UMI
Grabette is directly inspired by Stanford University’s Universal Manipulation Interface (UMI). This is a handheld gripper with a fisheye camera that records “real world” demonstrations, recovers camera trajectories with SLAM, and trains visuomotor policies from there.
UMI has proven that the recipe works. Other (closed source) devices also exist, such as Agibot’s MEgo gripper, Genrobot’s DAS gripper, and the Sunday Robotics skill capture glove. Our goal was to make it easy to use and to make the barrier between “I have a task” and “I have a trained model” as low as possible.
Grabette is built into a modern open ecosystem. That is, LeRobot for datasets, Hugging Face Hub for sharing, and a processing pipeline that runs from your browser without installing anything. Grabette is something anyone can build on their workbench, use in the field, and contribute data to.
Introduction to Gravette
We’ve been developing Grabette for a few months now, and we feel like it’s usable enough to share. I’m excited to share it with you right now.
Grabette is a handheld gripper that comes equipped with everything you need to rebuild your operational demonstrations.
It has two cameras, each with a different role. The separation of the two roles is intentional. An inexpensive wide fisheye provides the context-rich wrist camera-style view you need for policy, and an RGBD camera does the heavy lifting with robust 6-DoF tracking.
Grabette also records data during tasks performed by the user, but relies on its robot counterpart to perform the learned movements after training. Introducing Gripette, Grabette’s robotic arm end effector twin.
This family shares the same hardware DNA.
Grabette, a handheld demonstration device (camera + IMU + gripper, BOM cost ~490 euros) Gripette, a motorized gripper that closes the loop on a real or simulated robot arm (camera + 2 servomotors, BOM cost ~120 euros)
Grabette handheld recording device (left) and Gripette robotic gripper (right)
Built for everyone
Everything is open source
Check out the repository!
Hardware: CAD and production file capture for Grabette and Gripette Services: On-device Raspberry Pi software Processing pipeline: Run locally or online via Hugging Face Space Example downstream stack: Stock LeRobot training + OpenArm evaluation for reference
component. There are no standard sensors, closed pipelines, or fork lock-in available for purchase. Raspberry Pi, standard Pi camera, off-the-shelf OAK-D depth camera, and magnetic encoder. Importantly, anyone can assemble it using just the parts you order.
The design does not depend on robots. None of the capture or data formats are arm-specific. The demonstration is stored as a camera-local 6-DoF Cartesian pose and gripper state, and the output is a standard LeRobot dataset on the Hugging Face Hub, so the same data can drive different robots and different learning methods. However, the arm requires a matching Grippet gripper.
From your hands to your dataset in two steps
This release enables anyone to quickly go from “I want to demonstrate a task” to “I have a dataset ready for training” without prior expertise.
1. Records
At the push of a button, data from the observation camera, tracking camera (color, depth, IMU), and gripper encoder joint values are recorded simultaneously using a single shared clock to ensure proper synchronization. Pressing the button again will stop the episode and the data will be saved locally on the Raspberry Pi.
2. Process directly in the browser
Open the Grabette dashboard in your browser. Select the episodes you want to add to your dataset and start post-processing with one click.
Episodes are uploaded to the HF hub Gravet Slam Space runs SLAM using the RTAB-MAP library to ensure the trajectories are correct (no loss of jumps or tracking) Episodes are converted to LeRobot format New datasets are uploaded to the space and the data for each episode can be viewed using the Le Robot visualizer
Now you are all ready to start training.
What can you do with your data? Here’s an example.
Datasets are only interesting if you want to train them on something. Therefore, I will ship a complete example to demonstrate the loop end-to-end. However, I want to be clear: this is not a product, but an example of what is possible. A release is a recording system. Grabette data works with any method that uses the LeRobot dataset.
This example uses 200 recorded demonstrations:
and:
Train the policy using the LeRobot stack. This is a spreading policy (ResNet18 + SpatialSoftmax encoder, DDIM scheduler, 6-D rotation action) that fits a single consumer GPU. When evaluated on an OpenArm 7-DoF arm with a Gripette gripper, driven via the gRPC API, we get the following results:
now it’s your turn
Data bottlenecks are not solved by one lab, but by communities recording demonstrations everywhere. Now we can help you build your dataset. Build a Grabette, record tasks that interest you, and share them on the hub. With every episode, the open dataset gets bigger and more diverse, and every contributor makes robot learning a little less gated behind expensive hardware.
what’s next
This release is just the beginning of the project. Gravette will continue to evolve. More products are already on the way (still in development), including Casquette, a head-mounted POV device that complements Grabette for egocentric capture. But the most important next steps aren’t just ours, they’re yours too. Start recording and let’s build your dataset together.
👉 Build a Grabette (GitHub) · Process data (HF Space) · Contribute to a dataset
Built using ❤️ by Pollen Robotics.
