Grabette: an open system for collecting robotics data without a robot

Robot learning has a well-known supply problem. Transformer-based VLA architectures, diffusion and flow-matching policies, even world models exist today, and the compute — GPUs — to train them is available. What is missing is real-world manipulation data in large quantities and broad variety. Classic collection via teleoperation of an expensive robot is cumbersome, requires dedicated hardware and does not scale. Grabette, a recently released open-source system, sidesteps the robot entirely: a human hand, a gripper, a camera and a way to reconstruct the hand's 6-DoF trajectory are enough.

The system builds directly on the UMI (Universal Manipulation Interface) from Stanford, which demonstrated that visual-motor policies can be trained from in-the-wild recordings using SLAM. Closed-source counterparts already operate on the same principle — Agibot's MEgo gripper, Genrobot's DAS gripper, Sunday Robotics' skill-capture glove. Grabette's goal is to lower the barrier from "I have a task" to "I have a trained model" as far as possible, until recording a demonstration becomes training data accessible to anyone who wants to contribute to an open, collaborative dataset.

The hardware uses off-the-shelf components that can be ordered independently, with no lock-in to proprietary gear: a Raspberry Pi board, a standard Pi camera, an OAK-D depth camera and magnetic encoders. Grabette carries two cameras with distinct roles. A cheap fisheye camera feeds the visual reasoning model peripheral context in the style of a wrist-mounted camera. An RGBD camera does the heavy lifting of precise spatial trajectory tracking. After training, the data is deployed through Gripette, a compatible robotic gripper that serves as the end-effector and shares the same "hardware DNA" with the handheld recording rig.

Data is stored in a robot-agnostic format: demonstrations are saved as the camera's local Cartesian pose in six degrees of freedom together with the gripper state. The processing pipeline runs in the browser with no installation required, and the final output is a standard LeRobot dataset hosted on Hugging Face Hub. This design lets the same data drive different robots and learning methods, provided they mount a compatible Gripette gripper. The developers say the system was developed over several months and has reached a usability level that justifies public release; the stated aim is to let anyone go from wanting to demonstrate a task to holding a training-ready dataset, with no prior experience required.