Pollen Robotics introduced Grabette on the Hugging Face Blog as an open, low-cost rig that captures robot-ready manipulation trajectories using only a human hand. The system addresses the persistent shortage of diverse real-world data that now constrains transformer-based VLAs and diffusion policies more than model architecture or GPU availability.
Grabette mounts two cameras on a handheld gripper: a wide fisheye lens for policy context and an RGBD unit paired with IMU for robust 6-DoF pose recovery via SLAM. Demonstrations are processed directly in the browser, producing standardized outputs compatible with LeRobot without local installation or teleoperation hardware.
The design explicitly builds on Stanford’s Universal Manipulation Interface. Where UMI proved that in-the-wild recordings can train visuomotor policies, Grabette adds browser-native tooling and Hugging Face Hub integration to encourage community contributions at scale.
A motorized counterpart called Gripette completes the loop. Built from the same hardware DNA at roughly 120€ BOM cost, Gripette mounts on a robot arm to execute the learned trajectories, allowing teams to close the data-to-deployment cycle with minimal custom engineering.
Compared with closed alternatives such as Agibot’s MEgo or Genrobot’s DAS gripper, Grabette prioritizes replicability and zero-robot data collection. Any contributor can assemble the device on a workbench and record tasks in unstructured environments, bypassing the logistical overhead of maintaining teleoperation rigs.
The commercial signal lies in the shift toward shared, continuously growing datasets rather than proprietary collection campaigns. Labs that previously competed on data volume can now pool recordings through the Hugging Face Hub, potentially accelerating policy generalization across tasks and environments.
For hardware builders the release underscores two concrete bets: first, that accessible capture devices will outpace centralized teleoperation fleets in dataset diversity; second, that open pipelines linking handheld recording to LeRobot training lower the threshold for new entrants to validate end-to-end manipulation stacks.
Near-term use of the platform will likely focus on household and light industrial tasks where rapid iteration on gripper trajectories matters more than high payload. Teams adopting Grabette can therefore prioritize software iteration and community data contributions over capital-intensive robot fleets.