Digital Event Horizon
Grabette, a groundbreaking open system for recording robot-manipulation data, has been released to address the current bottleneck in the field of robot learning. With its user-friendly design and collaborative dataset goals, Grabette is poised to revolutionize the way we approach robot learning research.
Grabette is an open system designed to record robot-manipulation data for collaborative dataset creation. The system aims to bridge the gap between model development and data availability in robot learning. Grabette uses the Universal Manipulation Interface (UMI) concept, a handheld gripper system that records demonstrations "in the wild." The device is designed to be user-friendly, affordable, and robot-agnostic for diverse dataset use. The project invites community contributions to develop a collaborative environment for recording and sharing demonstrations.
Grabette is an open system designed to record robot-manipulation data, with the ultimate goal of creating a large, collaborative dataset for robot learning. This innovative tool aims to bridge the gap between model development and data availability, providing a solution to the current bottleneck in the field.
The development of Grabette is rooted in the concept of Universal Manipulation Interface (UMI), a handheld gripper system that records demonstrations "in the wild," recovers camera trajectories with SLAM, and trains visuomotor policies from them. The UMI proved that the recipe for successful robot learning works, but other closed-source devices exist, making it difficult to access high-quality data.
The creators of Grabette are addressing this challenge by building an open system that is accessible to everyone. This handheld gripper instrumented with cameras and sensors can record manipulation demonstrations in a user-friendly manner, using a two-step process:
1. Record: Users press a button, capturing data from the observation camera, tracking camera, and gripper's encoder joint values.
2. Process: The recorded data is uploaded to the Hugging Face Hub, where it undergoes processing, including SLAM verification and conversion to LeRobot format.
Grabette is designed to be robot-agnostic, allowing users to record demonstrations without a particular arm in mind. This means that the same dataset can be used with different robots and learning methods. The device itself is also built with affordability in mind, using standard sensors and components that can be easily ordered.
The creators of Grabette are now releasing the system, inviting the community to contribute to the development of this open dataset. By building upon the UMI concept, Grabette aims to create a collaborative environment where anyone can record demonstrations and share them on the Hub. This will lead to a more diverse and comprehensive dataset, ultimately benefiting robot learning research.
Casquette, a head-mounted POV device, is also in development to complement Grabette for egocentric capture. While this feature is still in progress, the release of Grabette marks an exciting step forward in the field of robot learning. The community's involvement is crucial in shaping the future of this project, and we can expect more updates and developments as time progresses.
Related Information:
https://www.digitaleventhorizon.com/articles/The-Dawn-of-a-New-Era-in-Robot-Learning-Introducing-Grabette-deh.shtml
https://huggingface.co/blog/grabette
https://github.com/pollen-robotics/grabette
Published: Tue Jul 21 04:57:20 2026 by llama3.2 3B Q4_K_M
