Robot learning has a supply problem. We have capable policy architectures and the GPUs to train them, but what we lack is large, diverse, real-world manipulation data. Teleoperating a robot to collect it is expensive and demanding, requiring a robot and hours of tedious work.
But what if you didn'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. That's the idea behind Grabette, an open, low-cost system for recording manipulation data.[reference:6]
Grabette is a handheld gripper instrumented with two cameras. A cheap wide fisheye provides a context-rich, wrist-camera-style view, while an RGBD camera does the heavy lifting of robust 6-DoF tracking.[reference:7] It records synchronized camera, IMU, and gripper encoder data, then uses SLAM to recover camera trajectories.[reference:8] The output is a standard LeRobot dataset on the Hugging Face Hub, ready for policy learning.
Grabette is directly inspired by Stanford's Universal Manipulation Interface (UMI), which proved that handheld grippers can record demonstrations "in the wild" and train visuomotor policies.[reference:9] But Grabette is built for everyone. Everything is open source: hardware CAD files, on-device capture software, and a processing pipeline that runs from your browser.[reference:10]
The release enables anyone to go from "I want to demonstrate a task" to "I have a training-ready dataset" quickly and without prior expertise. Press a button to record an episode, press again to stop, and the data is saved locally. Open the Grabette dashboard, select episodes, and with one click, post-processing begins. The pipeline runs SLAM using RTAB-MAP, verifies the trajectory, and uploads a new dataset to your Hugging Face Space.
To show the loop end-to-end, Grabette ships with an example downstream stack: a Diffusion Policy (ResNet18 + SpatialSoftmax encoder) that fits on a single consumer GPU, deployable on an OpenArm 7-DoF arm with the Gripette gripper.
The data bottleneck doesn't get solved by one lab, but by a community recording demonstrations everywhere. Grabette is a step toward seeding a large, open, collaborative manipulation dataset. Build a Grabette, record tasks you care about, and share them on the Hub. Every episode makes the open dataset bigger and more diverse, and every contributor makes robot learning a little less gated behind expensive hardware.
This release is only the start. Future plans include a head-mounted POV device for egocentric capture. But the most important next step is yours: start recording, and let's build the dataset together.