Hugging Face has released Grabette, an open system meant to make robot manipulation data easier to collect. Instead of requiring a robot, a lab, or a teleoperation rig, Grabette uses a handheld gripper with cameras to record how a person performs a task.
Robot-learning systems need diverse real-world demonstrations, but collecting them with physical robots is slow and expensive. Grabette is designed to capture a human hand’s six-degree-of-freedom motion, process the recording, and turn it into a robot-ready dataset.
The project is inspired by Stanford’s Universal Manipulation Interface. Hugging Face says Grabette works with LeRobot for datasets, the Hugging Face Hub for sharing, and a browser-based processing pipeline. The goal is not only a device but a shared, open dataset that many contributors can expand.
The system still depends on a robot counterpart to execute learned movements after training, so it does not remove robotics hardware from the full loop. It lowers the barrier at the data-collection stage, where robotics research often runs out of scale.