Hugging Face Ships Grabette for Open VLA Data Collection

On July 21, 2026, Hugging Face and Pollen Robotics released Grabette, an open-source handheld gripper built to record physical demonstration data. The system allows developers to step away from laboratory-bound teleoperation rigs and capture high-quality manipulation trajectories in any environment. This release directly addresses the physical data bottleneck constraining the training of vision-language-action models.

Hardware Architecture

Grabette functions purely as a data collection instrument. It is designed to pair with Gripette, its robotic end-effector counterpart used for actual policy execution. By sharing identical hardware specifications across both the collection and execution devices, the system minimizes transfer errors when deploying trained models to physical hardware.

The device relies on accessible, off-the-shelf components to ensure teams can replicate it at a low cost. The compute layer runs on a Raspberry Pi Zero 2W, balancing power efficiency with enough processing capability to manage the sensor streams. Two Feetech STS3215 servos provide actuation, utilizing magnetic encoders to capture precise gripper states during operation. These specific servos ensure the torque and resistance feedback roughly match the physical reality the executing robot will encounter.

Sensory input relies on a dual-camera configuration. A wide fisheye lens capturing video at 1296x972 provides a wrist-mounted perspective. An OAK-D depth camera handles the spatial recording, generating the necessary data for 6-DoF trajectory reconstruction.

Data Processing and Export

The hardware stack records synchronized visual and kinematic data at frequencies between 10Hz and 20Hz. The system uses Simultaneous Localization and Mapping (SLAM) algorithms to map the 3D trajectories of the hand as the operator physically guides the gripper through a task.

Data flows through a gRPC motor and camera service, which streams frames and positions over the network. Users can then clean and reconstruct these trajectories directly within a browser-based pipeline.

Processed datasets export natively to the standardized LeRobot format, ensuring immediate compatibility with Hugging Face’s broader open-source robotics ecosystem.

Open Infrastructure Strategy

The system architecture follows the design principles established by Stanford’s Universal Manipulation Interface (UMI). By standardizing the physical collection tool, Hugging Face aims to crowdsource a massive, collaborative data commons for robotic learning.

This hardware release builds on Hugging Face’s acquisition of Pollen Robotics in April 2025. The integration signals a clear operational mandate to establish the company as a primary hub for physical AI datasets, mirroring its position in text and vision models. Pollen Robotics also confirmed development of Casquette, an upcoming head-mounted device designed to capture egocentric point-of-view demonstrations.

If your team trains manipulation policies, Grabette offers a standardized path to build custom datasets without investing in stationary teleoperation infrastructure. Adopting the LeRobot format ensures your recorded trajectories remain compatible with standard open-source training pipelines.