ETH Zurich's Cableless Robot Hand Walks on Its Fingers

Could a robot hand abandon the arm and walk on its own fingers? Researchers at ETH Zurich have built an 818-gram cableless robot hand that does exactly that.

Using its five fingers as legs, the hand can crawl across different surfaces, recover from falls, and even interact with objects and a keyboard. The project offers a different approach to robot design: instead of adding more hardware, it aims to make existing hardware do more.

By combining a humanoid hand with reinforcement learning, the researchers turned a familiar robotic component into both a mobile platform and a manipulator. The work also reconsiders where the functional boundaries of a robotic part should lie.

Rather than treating a hand only as an end-effector for grasping, the ETH Zurich team explores whether the same structure can support locomotion. That shift in perspective is central to the project's contribution. At 818 grams, the hand is light enough to be tested as a standalone system, and its cableless operation removes tethers that would otherwise limit movement. The researchers' focus is not on a single application but on a design question: how much capability can be packed into one familiar robotic part?

Fingers as legs: one hand for mobility and manipulation

The hand's distinctive feature is that it handles both locomotion and manipulation at the same time. In conventional designs, a robotic hand is fixed to the end of an arm and is responsible for grasping and manipulation, while movement is handled by a base or legs. ETH Zurich's approach merges these two functions into a single hand: the fingers can grasp objects and also push against the ground to crawl. This reflects a deeper design principle: reduce the number of hardware components and let each part take on more functions.

Reinforcement learning teaches the hand to walk

Walking on five fingers is much harder to control than walking on two legs. The fingers have many joints, high degrees of freedom, and scattered contact points. Coordinating them into stable locomotion is difficult for traditional rule-based programming, which struggles to cover all situations. Small changes in contact or posture can require very different control responses, making hand-written rules brittle.

The researchers used reinforcement learning, letting the hand try repeatedly in simulation and in the real world. Through this process, it learned to support itself on its fingertips and move them in alternation. The cableless design means the hand does not rely on external power or communication, so it can act independently in real environments. That independence is important for testing whether the learned behaviors hold up outside a controlled laboratory setup.