Grabette: an open system to record robot-manipulation data

Details

Robot learning has a data supply problem: capable policy architectures exist, GPUs are available, but large-scale diverse real-world manipulation data is scarce. Teleoperating a robot to collect it requires owning a robot, is slow, and is difficult to scale across the task and environment variety modern policies need.\n\nGrabette is Pollen Robotics' answer: a low-cost, open-source handheld gripper with a camera that records 6-DoF hand demonstrations without a robot present. Users pick it up, perform a manipulation task, and get back a LeRobot-compatible dataset that can train a visuomotor policy. The processing pipeline runs from a browser with nothing to install.\n\nThe design is inspired by Stanford's Universal Manipulation Interface (UMI), which proved the handheld-with-fisheye-camera recipe works for in-the-wild data collection. Unlike UMI (and closed competitors like Agibot's MEgo and Sunday Robotics' skill capture glove), Grabette is fully open source and built into the Hugging Face Hub for dataset sharing.\n\nThe strategic ambition extends beyond the device: Pollen Robotics wants Grabette to seed a large collaborative manipulation dataset that no single lab could collect alone. If recording a demonstration becomes as easy as shooting a video, the contributor pool expands from robotics labs to anyone with a Grabette.