Swiss scientists turn a severed robotic hand into a walking creature that plays games

Engineers at the Soft Robotics Lab at ETH Zurich took an off-the-shelf five-fingered robotic hand called WUJI, crammed a tiny Raspberry Pi Zero 2 W, a battery, and an inertial sensor onto its back, and decided it didn't need an arm anymore. Weighing just 818 grams, the contraption looks like a high-tech nightmare, but the team successfully trained the disembodied limb to walk entirely on its fingertips.

Instead of manually programming an awkward spider-walk, the researchers threw the hand into Nvidia's Isaac Lab simulation to figure things out through reinforcement learning. Because human-like hands are notoriously asymmetrical—fingers have wildly different lengths and the thumb sticks out sideways—scripting balance by hand would have been agony. The algorithm had to learn from scratch which fingers to shuffle forward and which to plant firmly on the ground without faceplanting.

The digital training paid off in the real world: the autonomous hand scuttled across 14 radically different surfaces, including polished tile, slick wood, gritty asphalt, metal grating, artificial turf, and even loose gravel. It clocks an average speed of 9.3 centimeters per second. More impressively, whenever the clumsy limb took a tumble, it managed to flip itself back upright in 21 out of 25 attempts, effectively shrugging off its own lack of a torso.

Beyond basic locomotion, the robotic hand mastered precise physical interaction with external tools. In lab trials tracked by an overhead camera, it pushed around miniature blocks and tapped arrow keys on a standard keyboard with surprising accuracy, successfully hitting 29 out of 32 commands to play the classic puzzle game Sokoban.

The engineering logic behind the madness is simple: why force a massive robot into a tight vent or a collapsed building when its wrist could simply pop off, skitter into the dark crevice like Thing from the Addams Family, do its job, and stroll back out. Humanity spent decades worrying about giant mechanical terminators marching down the streets, completely missing the fact that future workplace surveillance might just crawl out of the ceiling tiles on five motorized knuckles.

Source: srl-ethz.github.io