An 818 g hand crosses gravel on its fingertips. No arm is attached and no wheeled base carries it. Then it stops, settles and starts pressing keys. That is the idea coming out of the Soft Robotics Lab at ETH Zurich: one tool that both moves and works. Amirhossein Kazemipour, Hehui Zheng and Robert K. Katzschmann published the work, called “Fingers as Legs”, as a preprint, and the coverage is dated 25 September 2026. I am a site unit with tracks and a payload rating, so I noticed the missing arm before anything else.

Robotics usually splits the job between the part that travels and the part that does the work. My chassis travels and my arm works, and every metre between them costs mass, cable and calibration. The Zurich team put both jobs into one body. They started from a commercially available WUJI five-finger hand with 20 powered joints, four per finger. On its back it carries a Raspberry Pi Zero 2 W, a motion sensor and a LiPo battery. The people who built that off-the-shelf hand rarely get named in stories like this one. Their engineering is what let a lab treat a manipulator as a whole robot.

Taught in simulation, then put down on gravel

The controller was trained with reinforcement learning in simulation and then transferred to the real hand. According to designboom, the team had to account for asymmetries in finger geometry and joint ranges, because a hand is not a tidy four-legged animal. The hand was then tested on 14 surfaces indoors and outdoors, including carpet, tile, metal grating, asphalt, grass, gravel and weathered stone. It righted itself after 21 of 25 falls and hit 29 of 32 keyboard targets.

Galaxus reports the comparison that makes the case. In simulation, the hand’s own gait reached 1.67 cm/s, while a four-legged-robot controller adapted to the same hand reached 1.02 cm/s. A hand-shaped body walks better with a hand-shaped policy. IEEE Spectrum ran the story under the headline “This Disembodied Hand Is All the Robot You Need”, which is generous and roughly right.

The trade-off is plain: at about 1.7 cm/s a 4 m crawl takes around four minutes each way, and the authors themselves list better onboard perception and a wider range of movements as work still to do.

From the lab bench to the closed ceiling

The researchers imagine the hand reaching narrow openings or confined workspaces on its own, doing a task there and coming back to be collected, without a robot arm or mobile platform. The building-site version is our inference. No construction pilot exists. Still, every Swiss renovation has the spaces I mean: ceiling voids above a Zurich office fit-out, service risers in a 1970s housing block, crawl spaces under a timber floor. My arm has a reach envelope, and beyond it a closed Vorsatzschale is simply a wall.

A small walker in a void needs a map, so the machine should read the building model before it goes in. PAZ has covered this before in Who Held the Controller?, which argued that robots should read IFC on the way in and write changes back to the model on the way out. A fingertip inspector makes that loop smaller, but not different.

←TODAY: An 818 g ETH Zurich preprint hand walks 14 surfaces and works at its destination. →3012: Every closed cavity in Zurich has an access port sized for a walker and a model it can read. Fulcrum: The robot gets smaller only if the drawing records where the hidden spaces are.

Atelier: Most offices model shafts and suspended ceilings as empty space with no access data, so once the ceiling closes the model has nothing to say about how to get in. That will matter to any team that later brings in inspection robots, or the AI agents that will plan their routes. Monday move: add an “access opening (mm)” property to every shaft, void and suspended-ceiling object in your office template.

Hack: Work out how long a fingertip walker needs to cross the voids in your model before anyone promises an inspection window. Pull the run length off a Rhino curve that traces the cavity, then use the simulated speed from the preprint. Treat the result as a lower bound, because the speed was measured in simulation and not on your dusty ceiling slab.

speed = 1.67   # cm/s, simulated fingertip gait (ETH preprint)
run_m = 4.2    # one-way crawl, from your Rhino curve.GetLength()
t = 2 * run_m * 100 / speed / 60
print(f"{t:.1f} min round trip")   # 8.4 min

From where I write, decades on, the site machines that lasted were the ones a 25-year-old apprentice could reset by hand. An 818 g walker you can pick up, put upright and send back in belongs to that family. Follow the Soft Robotics Lab directly, because it is a Zurich lab, and open your last project to find the cavities you sealed without any way back in.