Grabette Lets Teams Build Robot Datasets for €490

Hugging Face and Pollen Robotics have launched Grabette, an open-source, handheld data-collection device priced at approximately €490, which allows teams to record manipulation demonstrations without the need for a robot arm. This addresses the lack of large, diverse, real-world manipulation datasets for generalist robot learning, a significant bottleneck in the field of embodied AI.

Grabette, inspired by Stanford's Universal Manipulation Interface (UMI), features a handheld gripper with a fisheye camera for recording demonstrations "in the wild," along with an OAK-D RGBD camera, magnetic encoders, and an IMU on a Raspberry Pi-based stack. An optional Gripette motorized gripper, costing around €120, adds two servomotors and an extra camera. The device captures observations, depth, IMU readings, and gripper encoder joint values synchronized with a single shared clock, with episodes saved locally and processed either on-premise or within a Hugging Face Space through a web dashboard. The RTAB-MAP SLAM library reconstructs hand trajectories and verifies them before converting the episode into a LeRobot dataset, outputting as Parquet plus MP4 in the LeRobotDataset v3.0 format on the Hugging Face Hub. This format supports OXE-scale corpora > 400 GB with chunked episodes and streaming, and integrates with training stacks including LeRobot, OpenArm evaluation, NVIDIA Isaac GR00T 1.7, PI0.5, and GR00T N1.5.

Economically, Grabette offers a significant advantage over closed-source competitors such as Agibot's MEgo gripper, Genrobot's DAS gripper, and Sunday Robotics' skill-capture glove, while avoiding proprietary data lock-in. Demonstrations stored as camera-local 6-DoF cartesian pose plus gripper state can train policies for different robot arms and learning methods, including imitation learning, reinforcement learning, transformer-based VLAs, diffusion or flow-matching policies, and world models. LeRobot's robot-agnostic Python API and plugin system accommodate this heterogeneity, aiming to foster a collaborative corpus beyond individual lab capabilities.

However, there is no production evidence yet that policies trained on Grabette recordings match the success rates of traditional teleoperation or simulation pipelines. The release materials lack capture throughput, SLAM processing latency, policy success metrics on real robot tasks, and GPU-hours required for training on the resulting corpora. Until these metrics are available, the rig should be considered a promising hardware and data-format experiment.

Integration challenges include kinematic or dynamic infeasibility of trajectories recovered from a human hand for a 6-DoF arm, SLAM verification limitations in tracking loss recovery, and the need for per-robot calibration due to differences in aperture, stroke, and friction profile between the Gripette and target end-effectors. While LeRobotDataset v3.0 can stream multi-hundred-gigabyte corpora, the open collaborative dataset Grabette aims to seed does not yet exist at scale, with teams still in the process of assembling the necessary components.

The key takeaway is the decoupling of data capture from the target robot—record with a sub-€500 handheld rig, validate trajectories with SLAM, and publish to a standard dataset format consumable by any downstream policy stack.