Self-Contained Robotic Hand Crawls Surfaces, Balances, and Presses Keys

Researchers at the ETH Zurich Soft Robotics Lab have engineered a fully autonomous, walking robotic hand capable of navigating diverse terrains, self-righting, and performing dexterous manipulation tasks independently of a supporting arm. Led by Ali Kazemi, the team addressed a fundamental limitation in prior mobile hand designs, which typically relied on symmetrical fingers to simplify locomotion, thereby compromising the human-like finger asymmetry and opposable thumbs required for complex object interaction. The innovation centers on adapting an off-the-shelf human-like robotic hand with unequal fingers. To achieve autonomy, the researchers integrated onboard power, motion sensors, and a miniature computer directly onto the device. The control system was developed using reinforcement learning within a high-fidelity simulator, specifically leveraging NVIDIA Isaac Sim to train the hand on balance, gait, and manipulation strategies. The model incorporates stance-calibrated control frames and evaluates support polygons formed by fingertip contacts to maintain stability. Experimental validation demonstrated robust performance across multiple metrics. The mobile hand traversed 14 distinct surfaces ranging from indoor carpet to outdoor gravel, grass, and metal grating. In self-recovery tests, where the unit was initially placed on its side, it successfully pushed itself upright and restored its balance in 21 of 25 trials. Dexterous capabilities were verified through interaction tasks. While balancing on four fingers, the robot lifted the fifth digit to press specific arrow keys, achieving a 29 out of 32 success rate. Additionally, the hand navigated to a small block, pushed it across a surface, and positioned it within two centimeters of a target mark in repeated trials. The research, detailed in a paper posted to the arXiv preprint server, indicates that imbuing robotic hands with independent mobility allows them to access confined workspaces without the need for a carrier arm to follow. This advancement points toward more versatile mobile robots capable of operating effectively in human-centric environments, opening new possibilities for automation in restricted or unstructured settings.