ETH Zurich's Soft Robotics Lab equipped a commercial Wuji Technology hand with an 80-gram backpack containing a battery, IMU, and Raspberry Pi Zero, then trained it to walk using reinforcement learning in simulation. The hand navigated 14 surfaces including metal grates and gravel at 9 cm/s, self-righted in 21 of 25 trials, and pushed objects to targets up to 40 cm away with 100% success across 15 attempts.

The technical challenge centered on the hand's asymmetry — unlike traditional quadrupeds, fingers have unequal lengths and an opposable thumb. The team introduced "virtual springs" in the reward function that penalize lateral finger movement more heavily than forward motion, keeping the hand in a crawling posture while allowing stepping. All four control models (locomotion, self-righting, button pressing, object pushing) fit on the onboard Raspberry Pi.

Lead researcher Amirhossein Kazemipour positions this as a prototype for "autonomous modular embodiments" — robotic components that switch roles dynamically rather than remaining fixed to a single function. Potential use cases include detaching from an arm to access confined spaces in industrial maintenance or search-and-rescue scenarios, or operating controls unreachable by a full manipulator.

Columbia's Matei Ciocarlie notes the work suggests a "new class of mobile manipulators that are simultaneously dexterous, able to traverse complex terrain, and able to manipulate payloads comparable in size to themselves." Tokyo's Hideki Shimobayashi flags durability as an open question, since walking loads differ significantly from grasping forces.