News from the IT Home on September 29th: Researchers from the Soft Robotics Laboratory at the Swiss Federal Institute of Technology Zurich ( ETH Zurich ) have created a robotic hand that can walk on its own using its fingertips.

According to IT, this robot hand is modified from a conventional humanoid robot hand available on the market. It has a total of five fingers, each with four joints. The researchers did not redesign the structure of the hand but instead directly utilized these existing finger components and taught it how to move using reinforcement learning algorithms in a simulation environment. A battery, an onboard computer built using a Raspberry Pi (Raspberry Pi), and motion sensors have been installed in the palm area, allowing the entire device to operate completely independently. The weight of the whole machine is 818 grams.

Teaching this hand to walk is a great challenge; the fingers of the robotic hand are of varying lengths, with the thumb arranged opposite to the other four fingers. Whenever any finger is lifted in preparation to take a step, the entire body immediately loses its balance.

This neural network was trained on NVIDIA's Isaac Lab (NVIDIA's Isaac Lab) simulation platform. The network had to decide for itself when to lift each finger; the training process aimed for a stable posture and forward movement as rewards, until the robot arm developed its own walking gait. During the testing phase, it managed to crawl on 14 different indoor and outdoor surfaces, including asphalt roads, tiled floors, grass, gravel, and broken stone slabs. Out of 25 attempts where it was knocked down, it was able to adjust its posture and stand up on its own 21 times.

The fingers used for crawling can also restore the operational functions of a normal hand. During the demonstration, the robotic arm approached the keyboard to press keys with a correct key-pressing rate of ninety percent; it also completed a level of a puzzle game ( Sokoban ) involving pushing boxes, and was able to push small cubes to designated target locations with a positioning accuracy of millimeter level.

The actual use scenario envisioned by the researchers is to send this robotic arm to a narrow space that a complete large robot cannot reach, have it crawl inside, perform operations on objects or controllers, and then return. Whether this concept can be realized still needs to be verified through subsequent research and development. However, the thought of a robotic arm that is detached from its body walking towards you is still a bit unsettling.

Relevant papers:

https :// arxiv.org / abs /2609.17172

Advertising Notice: The external links contained within this text (including but not limited to hyperlinks, QR codes, passwords, etc.) are used to provide more information and save time in selection. The results are for reference only. All articles on IT that contain external links are subject to this notice.