Animals know that they need to keep on their toes at all times; otherwise, they could wind up as prey. As a result, some species have evolved to where newborns are capable of walking just minutes after birth- an astonishing feat engineers have been trying to adapt for robots. A research team from the USC’s Viterbi School of Engineering have taken a page from nature and developed a bio-inspired robotic limb that’s driven by animal-like tendons and powered by AI. The limb can be tripped-up and then recover before taking its next step.
The secret lies in a bio-inspired AI algorithm the team designed, which can learn new walking tasks on its own after as little as 5-minutes of unstructured play. For that to work, they used a model-free, open-loop approach that allows few-shot autonomous learning capable of producing active movements in a three-tendon, two-joint limb.
The limb itself was designed using a series of motors and joints to drive the tendons (3X tendons, 2X joints). One motor drives a tendon counterclockwise, while two others rotate a pair of tendons both clockwise and counterclockwise, and fourth actuates both clockwise, giving the leg freedom of motion similar to an animal.
The AI (known as General to Particular) learns how to move the limb(s) and maintain stability by undergoing a short period of motor babbling, which is then used to create an inverse map and followed by building functional habits, which are then reinforced with reward behavior with refinements to the map within a specific environment. Simply put, the AI will drive a limb in a particular motion in any given environment, if that motion is effective at moving the limb, it will remember what it did and use it again if it encounters that situation again.
The engineers state that their bio-inspired limb has potential applications in assistive technology, including robotic limbs and exoskeletons for people with mobility problems and those that have lost limbs. Search and rescue robots are on the list as well, allowing them to go into unfamiliar terrain unassisted.