機器人資料與工具
Grabette Open-Sources a Low-Cost Robot Demonstration Collector That Converts Hand Manipulation into LeRobot Datasets
Pollen Robotics has released Grabette, an approximately €490 handheld device that records six-degree-of-freedom manipulation using two cameras, an IMU, and a gripper encoder. Its browser-based workflow runs SLAM, validates trajectories, and exports standard LeRobot data.
Pollen Robotics has open-sourced Grabette in an effort to lower the barrier to collecting real-world robot manipulation data. Traditional demonstration datasets often require robotic arms, teleoperation equipment, and laboratory integration. Grabette instead uses a hand-operated gripper, a Raspberry Pi, a fisheye camera, an OAK-D RGBD camera, an IMU, and a magnetic encoder to synchronously record visual, depth, inertial, and gripper-state data. According to the project's estimate, the handheld device costs about €490 in materials, while the corresponding robot-mounted Gripette motorized gripper costs about €120.
The two cameras have clearly defined roles: the inexpensive fisheye camera provides a wide-angle view similar to that of a wrist-mounted camera, while the RGBD camera handles more robust six-degree-of-freedom tracking. After recording, users can select an episode in the browser. The data is uploaded to the Hub, where RTAB-MAP performs SLAM, checks the trajectory for jumps or tracking loss, and then converts it to the LeRobot format. The data is stored as a six-degree-of-freedom Cartesian pose in the camera's local coordinate frame, together with the gripper state, without being tied to a specific robotic arm.
The team trained a Diffusion Policy—small enough to run on a single consumer GPU—using 200 cup-grasping demonstrations, and demonstrated closed-loop execution on a seven-degree-of-freedom OpenArm equipped with Gripette. However, this is only a proof of feasibility and does not establish that the device is suitable for precision, contact-intensive, or high-speed tasks. SLAM drift, time-synchronization errors, camera occlusion, and differences between human and robot kinematics could all propagate into the training data. Its long-term value will depend on reproducible hardware calibration, data-quality labeling, and whether community-contributed data can transfer reliably across different robotic arms.