# ETH Zurich Engineers Build Autonomous Walking Robotic Hand for Confined Spaces Researchers turn a commercial robotic hand into a self-contained crawling manipulator. - **Source:** [IEEE Spectrum](https://spectrum.ieee.org/) - **Published:** 2026-10-08 - **Topic:** [Software & Engineering](https://brifing.si/c/engineering) - **Original article:** - **On Brifing:** https://brifing.si/s/40cfd960-5197-4e23-b0ca-509133a78a91 ## Autonomous Robotic Locomotion ETH Zurich researchers transformed an off-the-shelf robotic hand into a fully autonomous, mobile device capable of walking on its fingertips. Led by Ph.D. candidate Amirhossein Kazemipour, the project challenges traditional robotics by proving that an end-effector can navigate confined industrial or search-and-rescue environments without requiring an entire robotic arm. ## Off-The-Shelf Hardware Architecture Instead of fabricating custom hardware, the engineering team used a commercial hand from Wuji Technology. They integrated an 80-gram backpack containing a battery, an inertial measurement unit, and a Raspberry Pi Zero. This approach preserved the standard grasping capability while adding the computing power needed for untethered, standalone locomotion. ## Asymmetry in Locomotion Training Training the hand to walk required overcoming structural asymmetry. Unlike symmetric quadruped robots, human-like hands have unequal digit lengths and an opposable thumb. The team used reinforcement learning in simulation, applying distinct penalties and rewards to manage the complex, off-center dynamics inherent to crawling with uneven mechanical fingers. ## Virtual Spring Control Formulation To maintain a functional crawling posture, researchers established target positions relative to the palm for each finger. They created virtual springs by penalizing deviations from reference points. Penalties for forward and backward steps were significantly lower than lateral movements, encouraging efficient forward gait dynamics without compromising balance. ## Embedded Multimodal Policy Execution The onboard Raspberry Pi Zero hosts four distinct machine learning models: crawling, self-righting after falls, keyboard typing, and object pushing. The system dynamically switches between these policies depending on operational requirements, executing complex control loops directly on low-power embedded hardware without reliance on external compute clusters. ## Empirical Terrain Navigation Results Real-world testing verified locomotion across 14 diverse surfaces, including gravel, grass, smooth floors, and metal grates. The hand achieved an average crawling speed of 9 centimeters per second, successfully turned left and right, and demonstrated self-righting capabilities in 21 out of 25 distinct experimental fall trials. ## Precision Manipulation Performance Using an overhead camera feed, the autonomous hand pushed a 41-gram cube toward specified targets up to 40 centimeters away across 15 consecutive trials. This demonstration confirmed that the appendage can simultaneously support its structural payload during transit and execute delicate physical manipulation tasks upon arriving at destination targets. ## Paradigm of Modular Embodiments This breakthrough advances what Kazemipour terms autonomous modular embodiments, where robotic components detach to act as independent agents. While external experts note challenges regarding mechanical wear and onboard perception, the project validates multi-role sub-assemblies, pointing toward flexible robotic architectures that dynamically shift between manipulation and locomotion. ## Sources covering this story - https://spectrum.ieee.org/walking-robotic-hand — --- _Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._