ETH Zurich robot hand crawls like 'Thing' from The Addams Family
Researchers at ETH Zurich are teaching a commercially available robotic hand an unusual dual function: its fingers serve both for locomotion and for manipulating and moving objects.
The «Thing» from the «Addams Family» walks on its fingers as naturally as if they were legs. The robotic hand from the Soft Robotics Lab, however, first had to learn these movements. It walks over grass, gravel, and other surfaces, rights itself after a fall, and can also operate a keyboard.
The research project is called «Fingers as Legs». Amirhossein Kazemipour, Hehui Zheng, and Robert Katzschmann have published their results as a scientific preprint. As such, other researchers have not yet independently peer-reviewed the study.
Five fingers replace legs
The base is a commercially available WUJI robotic hand. Its five fingers have 20 powered joints. On the back of the hand, there is also a Raspberry Pi Zero 2 W, a motion sensor, and a lithium-polymer battery. The complete system weighs 818 grams.
Source: ETH Zurich / Soft Robotics Lab
The shape of the hand makes walking challenging. The fingers differ in length and position. The thumb is also offset to the side. Four-legged robots have it easier here because their legs are usually arranged much more symmetrically.
The researchers therefore trained the control system using reinforcement learning. In this process, a system tries out many movements and receives feedback on how well they meet the specified goal. The training initially took place in a simulation. There, the neural network learned for itself when to move which finger.
The control system rewards movements in the desired direction and a stable posture. In the simulation, the hand reached a speed of 1.67 centimeters per second. An adapted control system for four-legged robots achieved 1.02 centimeters per second under the same conditions.
From the lab to gravel and asphalt
On the actual hardware, the team tested the hand on 14 different surfaces. These include grass, gravel, asphalt, carpet, tile, hardwood, concrete, and metal grates. In addition, the researchers tested how well the hand rights itself after a fall. A specially trained control system brought it back onto its fingertips in 21 out of 25 attempts.
Source: ETH Zurich / Soft Robotics Lab
However, no single model currently controls all capabilities. The researchers use different control systems for walking, standing up, and other tasks. Which task the hand performs is specified from the outside. Changing direction also still poses difficulties for the asymmetrical hand.
At the destination, the legs become fingers again
After walking, the fingers change their task. One moment they are carrying the hand across the floor, and the next they are gripping, pushing, or moving objects.
The team demonstrates this on a keyboard, among other things. The hand presses specified keys one after the other. It works without a camera. Out of 32 commands, it executed 29 correctly. Before the attempt, however, the researchers had to position it appropriately in front of the keyboard.
In another experiment, the robotic hand pushed a cube to various target points. A camera above the experimental setup provided the position data for this. After 15 runs, the cube was on average 17 millimeters away from the specified target.
In the long term, this principle could make robots more flexible. The researchers describe, for example, a robotic arm that places its hand in a hard-to-reach location. From there, it could independently crawl to a switch or object, perform its task, and then return.
My interests are varied, I just like to enjoy life. Always on the lookout for news about darts, gaming, films and series.
From the latest iPhone to the return of 80s fashion. The editorial team will help you make sense of it all.
Show all