Grabette: an open system to record robot-manipulation data

Source: HuggingFace Date: 2026-07-21 URL: https://huggingface.co/blog/grabette

Summary

Grabette is an open-source ~€490 handheld gripper (Raspberry Pi, camera pair, OAK-D depth sensor, IMU, magnetic encoders) that lets anyone record human manipulation demonstrations without a robot or teleoperation rig. Recordings upload to a browser dashboard that runs SLAM validation, converts to LeRobot format, and publishes to Hugging Face Hub; the team demonstrated a diffusion policy trained on 200 cup-grasping demos deployed successfully on a 7-DoF robotic arm. A robot-agnostic companion device (Gripette, ~€120) exists too.

Implications

Feeds the physical-AI/robotics thread from the data side rather than the model side — robot learning’s actual bottleneck is diverse real-world manipulation data, and Grabette lowers the collection barrier from expensive robotic teleoperation setups to a €490 handheld device anyone can use. Complements today’s Cosmos 3 Edge signal (cheap open edge inference + cheap open data collection is the same de-centralization move); the LeRobot format target also suggests convergence toward a shared open robotics-data standard rather than fragmented per-lab formats. Worth watching whether a community dataset actually accumulates or this stays a hardware curiosity.