AIBRIEF
The handheld system, built with Pollen Robotics, records manipulation demos with a fisheye camera and motion tracking, no robot arm or teleoperation rig required.
Robot learning has plenty of capable policy architectures and not nearly enough demonstrations to train them on. Hugging Face and Pollen Robotics released Grabette on July 21, 2026 as a low-cost, open bid to fix that supply gap.
A person picks up a small handheld gripper, performs a manipulation task with their own hand, and a fisheye camera mounted on the device uses SLAM (motion tracking that recovers the camera's path from video) to reconstruct the gripper's trajectory as it moves. The recording pipeline runs in a browser; no robot arm, teleoperation rig, or local install is required. Datasets land on the Hugging Face Hub in LeRobot, the company's open robotics framework.
The post names Stanford's UMI (Universal Manipulation Interface) as the direct inspiration, and frames the project against closed alternatives: Agibot's MEgo gripper, Genrobot's DAS gripper, and Sunday Robotics' skill-capture glove. The bet is that lowering the floor for recording will seed a community dataset no single lab can build alone.
Code and a companion data repo are public at pollen-robotics/grabette and pollen-robotics/grabette-data. Early uptake shows on the Hub: user-contributed datasets like glannuzel/grabette-dataset-2 and SteveNguyen/grabette-dataset-cup-0107 are already posted.
Whether hobbyist-scale contributions reach a size that actually moves policy quality remains the open question. The release is the bet, not the proof.