Science1 publisher2 min readPublished
ETH Zurich engineers teach a stock humanoid hand to walk on its uneven fingertips
ETH Zurich engineers trained an off-the-shelf, 20-joint robot hand to walk on its uneven fingertips across 14 indoor and outdoor environments. Earlier walking hands only balanced after their fingers were rebuilt to one length, while this one keeps the thumb and uneven fingers that make a hand good with tools.
The Scientist · Science desk
Drafted by a language model from the sources cited here and checked against its claim ledger before publication. How we use AISend a correction
What happened
- The hand comes from ETH Zurich's Soft Robotics Lab and is described in a preprint that has not yet been peer reviewed.
- A battery, motion sensors and a small computer sit on the back of the hand near the wrist, and the whole unit weighs 1.8 pounds.
- While balancing on some fingers, the hand used the others to push a small cube forward.
- It also played the puzzle game Sokoban, pressing the right arrow keys in sequence.
- After a fall it can right itself, lifting itself up from lying flat.
Compiled by The ScientistSomething wrong?How this is made
Why it matters
- capability If the per-finger timing holds up, designers would not have to give up the thumb-and-uneven-finger layout used for tools to get a hand that walks.
- constraint The typing score rests on 32 keystrokes, and its plausible accuracy spans roughly 76% to 97%, too wide a band to plan a keyboard or control-panel task around.
- decision Humanoid builders weighing a detachable, crawling hand are looking at a proposed use; the tests described ran the hand on its own, never fitted to a robot.
A hand shaped like a person's is a poor walker. Its fingers differ in length and its thumb sits opposite the rest, Popular Science reported. That asymmetry partly explains human skill with tools, and it suits upright walking badly [3]. Earlier walking robot hands dealt with it by rebuilding the hand so every finger was the same length [2]. The ETH Zurich group kept the uneven shape of a stock hand sold for humanoid robots and left the controller to cope with it [4].
The controller was trained in simulation with reinforcement learning. The software earned a reward each time the hand balanced correctly on its uneven fingers [6]. To fit the odd shape, training also taught the software each finger's own timing [7]. Every finger got its own target position, so where and when it moved changed with the task [7].
Out of simulation, the hand crossed carpet, gravel, grass and metal with what the magazine called a springy, spider-like gait [8]. Balanced on some fingers, it typed with the others and hit the correct key 29 times in 32 [10]. That is about 91% [1]. The thing this doesn't tell you is the error rate over a long session. A standard 95% confidence interval on 29 successes in 32 tries runs from roughly 76% to 97% [2]. The account does not report walking speed, battery life, or how many trials were run in each environment.
The use the team proposes is a humanoid robot in a cramped space that reaches toward a spot it cannot get to, detaches its hand and sends it crawling off to fetch an item [13]. "Giving robotic hands their own mobility could make future robots more versatile in the spaces and interfaces built for people," the researchers wrote [14]. I think the nearer payoff is the second one they name, because it needs only the per-finger control and no hand that leaves the arm: refinements in finger articulation and balance that could make robots better at handling objects designed for human hands [15].
The humanoids such a hand would serve are also early. Popular Science notes they often still need human operators running them remotely [16].
What to watch
- Whether the preprint passes peer review, and whether the published version reports trial counts per environment and how often the hand fell.
- A demonstration of a humanoid robot detaching the hand, sending it to fetch an object, and reattaching it.
- Larger typing or manipulation trials that would narrow the roughly 76% to 97% range on keystroke accuracy.