Why CLAVIR

Dexterous manipulation remains one of the most difficult problems in robotics. Despite rapid progress in machine learning and robot perception, many everyday manipulation tasks remain surprisingly hard for machines.

A central reason is simple: most robotic hands were not designed for learning.

Traditionally, robotic hands have been engineered as precise actuators. They are expected to execute commands accurately and repeat motions reliably. While this works well for structured industrial tasks, it becomes limiting when robots must interact with complex, unstructured environments.

Manipulation in the real world is fundamentally contact-rich. Objects slide, roll, pivot, and collide. Contacts appear and disappear dynamically, and the robot must continuously adapt to these changes.

For learning systems, these interactions are not just disturbances — they are signals.

At CLAVIR, we start from a simple view: the hand should not merely execute actions, but help reveal the structure of interaction.

A robotic hand designed for learning should: expose meaningful physical signals, stabilize interaction with objects, and allow learning systems to explore rich control behaviors. In other words, the hand itself becomes part of the learning system.

This perspective suggests a different approach to designing dexterous robotic hands. Instead of treating hardware as a fixed interface beneath software, we explore how the physical structure of the hand can directly support learning.

CLAVIR is an ongoing effort to investigate this idea. Our goal is to build robotic hands whose physical properties improve the stability, observability, and expressiveness of contact-rich manipulation.

The CLAVIR Hand is currently under development. More details will be shared as the project evolves.