Researchers at ETH Zurich’s Soft Robotics Lab have built a disembodied robotic hand that can walk, balance and recover from falls using only its fingers — and the team openly compares it to Thing, the autonomous hand from The Addams Family.
The lab released videos and a pre-print paper describing the project, which it calls a demonstration of a self-contained, anthropomorphic hand capable of locomotion without any separate wheels, legs or support structure.
Amirhossein Kazemipour, a doctoral candidate at the Soft Robotics Lab, said teaching a robotic hand to walk unaided is harder than it looks.
“Getting a robotic hand to walk is actually pretty challenging,” he said in a video released by the school. “You know, it’s fingers aren’t all the same, so it has to learn how to coordinate all of them while remaining balanced.”
The hand faces an unusual engineering problem: it has to use the same fingers to move its own body, bear its own weight, and interact with objects around it. According to the pre-print paper, the hand retains its original finger design and position controller while learning these skills, and carries its own power and computing onboard, making it a fully self-contained platform.
The team used a reinforcement learning approach that accounts for the fact that the hand’s fingers are not identical in length or structure. Training took place in a simulator calibrated against measurements from the physical hardware. The paper states that, under the researchers’ reward formulation, the simulated hand moved faster than it did using reward systems originally designed for four-legged robots.
On the physical hardware, the team reports that task-specific control policies allowed the hand to crawl untethered, steer itself, and recover after falling over. While supporting its own weight, the hand was also able to carry out a sequence of keyboard commands without using vision, and to push an object toward target locations using overhead visual feedback.
The researchers describe the result as a compact mobile manipulator that reuses its fingers for both movement and interaction with its surroundings, rather than relying on a separate mechanism for locomotion.
The project underscores a broader challenge in humanoid robotics: replicating the dexterity of the human hand, which contains 27 bones — eight in the wrist, five in the palm, and fourteen in the fingers and thumb — along with 34 muscles controlling both fine and broad movements. Matching that range of motion and control remains one of the toughest problems for engineers working toward android-like robots.
