Grabette: an open system to record robot-manipulation data

Pollen Robotics Releases Open-Source Grabette to Crowdsource Robot Data

Pollen Robotics has released Grabette, an open-source, handheld system for recording robot manipulation data. The project aims to solve a critical bottleneck in robot learning: the lack of large, diverse, real-world datasets. By allowing anyone to capture complex tasks using a low-cost gripper instead of a full teleoperation rig, the initiative seeks to build a collaborative, open dataset that could significantly accelerate progress in training capable robot policies.

How Grabette Works

Inspired by Stanford's Universal Manipulation Interface (UMI), Grabette integrates off-the-shelf components to capture high-quality demonstration data. A wide fisheye camera provides the context-rich view for the policy, while an OAK-D RGBD camera and IMU handle the 6-DoF tracking via SLAM. The system is designed as an end-to-end open ecosystem, integrating with popular platforms and standards in the AI community.

• Hardware: The handheld Grabette has a bill of materials (BOM) cost of ~€490, while its robotic counterpart, the Gripette end-effector, costs ~€120. All CAD and production files are open source.
• Data Pipeline: Raw data is uploaded to a Hugging Face Space where it is automatically processed using the RTAB-MAP library, verified, and converted into the LeRobot dataset format.
• Robot-Agnostic: The output is a standard camera-local 6-DoF Cartesian pose, making the data usable across different robot arms, provided they are equipped with the Gripette gripper.

Impact on the AI Ecosystem

By lowering the barrier to entry for data collection, Grabette challenges the traditional, capital-intensive model of robotics research. This open, community-driven approach offers a stark contrast to closed-source commercial systems from companies like Agibot and Genrobot. If the community adopts the platform, the resulting large-scale dataset could become a foundational resource for training more generalist visuomotor policies. Pollen Robotics also teased future developments, including a head-mounted device named Casquette for egocentric data capture.

The strategic play here is not just the open-source hardware, but the frictionless, browser-based pipeline that connects individual contributors directly to the Hugging Face Hub. By standardizing on the LeRobot format, Pollen Robotics is building a data collection engine for the entire open-source robotics community, creating a network effect that isolated, proprietary systems cannot easily match.