ETH Zurich Robotic Hand Walks on Fingertips, Plays Sokoban

At ETH Zurich’s Soft Robotics Lab, scientists have developed a five-fingered robotic hand capable of walking on its fingertips, steering, recovering from falls, and manipulating objects autonomously. This hand, derived from a commercially available anthropomorphic model featuring 20 actuated joints (four per finger), weighs 818 grams and includes an onboard battery, sensors, and computing hardware. The research team utilized reinforcement learning to train locomotion policies in a simulated environment, which were then applied to the physical hand. The robot adeptly traversed 14 various surfaces and successfully righted itself in 21 out of 25 trials when tipped over. Additionally, it pressed keyboard arrow keys accurately 29 out of 32 times and completed tasks in the puzzle game Sokoban. Led by Amirhossein Kazemipour, Hehui Zheng, and Robert K. Katzschmann, the project seeks to integrate locomotion and manipulation into one system, potentially allowing the hand to operate in confined spaces. Future efforts will focus on enhancing onboard perception and broadening movement capabilities.