Grabette: an open system to record robot-manipulation data

Robot learning has never really had a model problem. It's had a data problem. Transformer-based VLAs, diffusion policies, world models — the architectures are there and so are the GPUs. What's missing is enough real-world manipulation footage to actually train them on, and teleoperating an actual robot arm to gather that footage is slow, expensive, and requires, well, a robot.

Pollen Robotics thinks it has a workaround, and it's called Grabette. It's a handheld gripper you carry around like a strange power tool, fitted with a Raspberry Pi, a standard Pi camera, an OAK-D depth camera, and magnetic encoders. You press a button, perform a task with your own hand — picking something up, stacking it, whatever — and press it again to stop. That's the whole recording process. No lab, no rig, no robot in sight. The bill of materials runs about 490 euros, which is pocket change compared to most teleoperation setups.

The idea isn't new on its own; it borrows directly from Stanford's Universal Manipulation Interface, which proved a few years back that a handheld gripper with a fisheye camera and SLAM tracking could recover usable 6-DoF trajectories from human demonstrations. Closed variants exist too, from Agibot's MEgo to Genrobot's DAS gripper. What Pollen is doing differently is making the whole pipeline open and frictionless: record on the device, upload to the Hugging Face Hub, run SLAM and format conversion through a browser-based dashboard, and out comes a standard LeRobot dataset ready to train on. Nothing to install, nothing robot-specific baked into the data format — it's stored as camera-local pose plus gripper state, so in theory any arm with the right end-effector can use it.

That end-effector is Gripette, Grabette's robotic twin — a motorized gripper, about 120 euros in parts, that mounts on a real arm and executes what the policy learned. Pollen demonstrates the full loop with 200 recorded episodes, training a diffusion policy (ResNet18 encoder, DDIM scheduler) that runs on a single consumer GPU, then evaluating it on an OpenArm 7-DoF arm over gRPC. They're careful to frame this as a proof of concept, not the actual product — the product is the recording system itself.

The bigger swing here is the ask: build a Grabette, record whatever tasks you care about, upload the episodes, and help grow a shared dataset that no single lab could assemble alone. It's a reasonable bet that if recording a demonstration is as easy as shooting a short video, a lot more people will actually do it. A head-mounted version called Casquette is apparently already in the works for egocentric capture.