HomeAIGrabette Released to Record Robot-Mani
AI

Grabette Released to Record Robot-Manipulation Data

Pollen Robotics released Grabette, a handheld device costing around 490 euros to record robot manipulation data without requiring a full robot arm.

WHAT YOU NEED TO KNOW
  • Pollen Robotics released Grabette, a handheld device costing roughly 490 euros to collect robot manipulation datasets.
  • The companion Gripette motorized gripper costs about 120 euros to execute trained actions on physical or simulated robot arms.
  • Data processing runs in a browser using the RTAB-MAP library for visual SLAM before publishing datasets to the Hugging Face Hub.
  • A reference pipeline trained a Diffusion Policy on 200 demonstrations using a single consumer GPU.

Pollen Robotics released Grabette, an open-source handheld gripper designed to capture robot manipulation data, according to an announcement published on Hugging Face. The system allows operators to record tasks using a handheld device and turn them automatically into datasets for robot learning.

Data collection for robot learning traditionally requires operating a physical robot arm, presenting hardware and logistical hurdles. Grabette bypasses the need for a robot during capture by using a human operator equipped with an instrumented handheld gripper, cameras, and trajectory tracking tools.

The handheld Grabette unit carries a bill-of-materials cost of approximately 490 euros. It combines two cameras: a wide fisheye camera providing context-rich wrist views and an off-the-shelf OAK-D RGBD depth camera for 6-degree-of-freedom tracking. The assembly incorporates a Raspberry Pi, a standard Pi camera, magnetic encoders, and an inertial measurement unit. Data streams from sensors and encoders record simultaneously using a single shared clock.

To execute learned movements on physical or simulated robots, Pollen Robotics designed Gripette as a motorized twin gripper. Gripette features a camera and two servomotors with a bill-of-materials cost of roughly 120 euros. Demonstrations store camera-local Cartesian poses and gripper states, outputting standardized LeRobot datasets hosted on the Hugging Face Hub.

Processing and Reference Implementation

Users process recorded episodes through a browser dashboard without installing local software. The pipeline uploads raw data to Hugging Face, where the grabette-slam space executes visual simultaneous localization and mapping using the RTAB-MAP library. The system verifies trajectory continuity before creating a formatted dataset.

Pollen Robotics published a reference workflow using 200 demonstrations to train a Diffusion Policy on a single consumer GPU. The training configuration uses a ResNet18 encoder with SpatialSoftmax and DDIM scheduling to predict 6-dimensional rotation actions. Testing was conducted on an OpenArm 7-degree-of-freedom arm fitted with the Gripette gripper over a gRPC connection.

The hardware CAD files, Raspberry Pi capture software, and processing pipelines are available under open-source licenses. Pollen Robotics also confirmed ongoing work on Casquette, a head-mounted point-of-view device for egocentric video capture. Following the release, the team made its electronic design repository public on GitHub after early users identified access limits.

Xentir Media
Xentir Media NewsroomSource-backed AI and technology coverage, drafted by Xentir's automated editorial system under fixed human-set rules. See our editorial policy and AI usage policy.
J
Jomon · Founder & EditorFounder and editor of Xentir Media. Sets the editorial rules the newsroom system runs under, and is accountable for its corrections. About Jomon · [email protected]
The Xentir Brief
The developments worth knowing — one useful email.
Get the Brief →