Grabette: an open system to record robot-manipulation data

Hugging Face has introduced Grabette, an open and low-cost system designed to record robot-manipulation data. According to the source, robot learning faces a data supply bottleneck where capable policy architectures and GPUs exist, but large, diverse, real-world manipulation data remains scarce. Traditional teleoperation requires expensive robots, labs, and specialized rigs, limiting scalability. Inspired by Stanford's Universal Manipulation Interface, Grabette allows users to record tasks using a human hand and a handheld gripper, bypassing the need for a physical robot during data collection. Built from standard, off-the-shelf components including a Raspberry Pi, a wide fisheye camera, an OAK-D depth camera, and magnetic encoders, Grabette captures observation data, tracking data, and gripper joint values synchronized via a single clock. Users can process episodes through a browser dashboard and automatically output standard LeRobot datasets on the Hugging Face Hub. While the capture format is robot-agnostic, executing trained movements requires Gripette, the robotic arm end-effector twin of Grabette. The release aims to seed a large, open, and collaborative manipulation dataset contributed to by a community of users worldwide, thereby reducing hardware-gated barriers in robot learning.