Grabette is an open, low-cost system from Pollen Robotics for recording robot manipulation data by hand, without requiring a robot or teleoperation rig.

The core insight is that the bottleneck in robot learning is lack of diverse real-world data, not model architecture.

Grabette, inspired by Stanford’s UMI, uses a handheld gripper with two cameras: a cheap wide fisheye for policy context and an RGBD camera for robust 6-DoF tracking via SLAM.

The companion Gripette is a motorized gripper that executes learned policies on a robot arm. All hardware (BOM ~€490 for Grabette, ~€120 for Gripette) and software are open source.

Recording is a two-step process: press a button to capture synchronized camera, IMU, and encoder data on a Raspberry Pi, then process episodes in a browser via a Hugging Face Space that runs SLAM (RTAB-MAP), verifies trajectories, and converts data to standard LeRobot datasets stored on the Hub.

The system is robot-agnostic—data is stored as camera-local 6-DoF pose plus gripper state.

An end-to-end example trains a Diffusion Policy on 200 demonstrations and evaluates it on an OpenArm 7-DoF arm with Gripette.

The stated, ongoing next steps include a head-mounted device called Casquette for egocentric capture.

The project’s goal is to seed a large open collaborative manipulation dataset that no single lab could build alone.