source: hugging face blog: grabette: an open system to record robot-manipulation data
level: technical
grabette is a handheld gripper with two cameras, an imu, and magnetic encoders. it records human demonstrations of manipulation tasks. one camera is a wide fisheye for policy training, the other is an rgbd camera for 6-dof tracking. the device costs about 490 euros in parts. a companion motorized gripper called gripette costs about 120 euros and mounts on a robot arm for execution. all hardware designs and software are open source.
to use grabette, press a button to start recording. data from both cameras, imu, and gripper encoder are saved with synchronized timestamps. press again to stop. then open the grabette dashboard in a browser. select episodes and click to process. the pipeline uploads data to hugging face hub, runs slam to verify trajectories, converts to lerobot format, and creates a dataset ready for training. no local installation is needed.
the system builds on the universal manipulation interface concept from stanford. it uses standard parts like a raspberry pi and off-the-shelf cameras. demonstrations are stored as camera-local 6-dof poses and gripper states, making them robot-agnostic. an example trained a diffusion policy on 200 demonstrations and deployed it on an openarm robot with gripette. the goal is to grow a shared, diverse manipulation dataset through community contributions.
why it matters: it lowers the barrier to collect real-world robot manipulation data, enabling faster progress in robot learning without expensive labs.
source: hugging face blog: grabette: an open system to record robot-manipulation data