A robotic hand is usually the part of the robot that gets carried around. Researchers at ETH Zurich have flipped that arrangement: they took the Wuji robotic hand, put it on the floor, and taught it to carry itself. It crawls, changes direction, and picks itself up after a fall, with no wheeled base or arm underneath.
The hand is the whole robot
The usual design splits the labour. A platform moves, an arm reaches, and a hand does the fiddly bit at the end. Here the fingers do all three jobs. They hold the body up, push it along, and, when needed, handle objects.
That makes the machine compact, which is the point. It also means the fingers have to be good at two very different things at once: locomotion, which wants steady and repeatable motion, and manipulation, which wants precision.
How it learned to move
The team used reinforcement learning, where a system improves by being scored on how well it performs a task. The training ran in a computer simulation, set up using measurements of the real device so the virtual hand resembled the physical one.
The algorithm also accounted for the fact that the fingers are not identical. They differ in length and in other characteristics, and a controller that pretends otherwise tends to learn a gait the hardware cannot perform.
The authors say their reward scheme produced faster movement in simulation than settings originally designed for four-legged robots. That is a reasonable thing to compare against, since a hand walking on its fingers has more in common with a quadruped than with a typical gripper. It is worth remembering, though, that this is a simulation result, and the report gives no speeds or figures.
What the physical tests showed
On real hardware, the hand moved without external power or external computing equipment. The energy source and the computing were built into the system itself. For a robot that is essentially a palm with fingers, that is a respectable amount of packing.
The researchers also tested whether the same body could do careful work. Two experiments stand out:
- Blind typing. The robot pressed keyboard keys one after another without using vision.
- Guided pushing. In a second test, it pushed an object to specified points, steering by the image from a camera mounted above.
The overhead camera is a small asterisk on the self-contained story. The report says the hand ran without external power and computing, but it does not say how the camera feed was handled in that experiment. The keyboard test, by contrast, needed no vision at all.
What we do not know
The report is brief, and several things are missing. It gives no crawling speeds, no success rates for the keyboard or pushing tasks, no count of fall recoveries, and no details about battery life or the onboard computer. It does not name a publication or paper either.
So the claim is best read as a demonstration, not a benchmark. One design walked across a surface, recovered from falls, pressed keys, and moved objects. Whether it does any of these quickly or reliably is not something the available material can say.
Why it is interesting anyway
The appeal is the consolidation. Robots tend to accumulate parts as their jobs expand, a base here, an arm there, a gripper on top. This project goes the other way and asks how much a single hand can do if it is allowed to stop being an accessory.
The answer, for now, is that it can walk, get up, type, and shove things around. Not bad for an appendage that used to wait for someone else to bring it to work.