A robotic leg without prior knowledge has learned to walk by relying on its artificial intelligence (AI) software. The robotic leg, which moves using animal-like tendons, can be tripped up but then recover before the next footfall. The leg was not programmed to do that, but taught itself.
Researchers at the USC Viterbi School of Engineering report the breakthrough in the journal Nature Machine Intelligence. [nature.com] The researchers developed a biology-inspired algorithm that learns new walking tasks by itself after 5 minutes of unstructured play. After that, the robot adapts to other tasks without more programming.
The robot learned to understand its environment in a process of free play, called 'motor babbling'. The robot monitored random movements of the leg to build an internal map of its limb and its interactions with the environment. By learning from experience, this kind of robot will figure out solutions to problems such as how to step over an obstacle without tripping over it. Pre-programmed robots eventually fail because programmers cannot predict and code for all possible scenarios a free moving robot might encounter.
Each robot's unique experience develops a walking gait works well enough, but it produces robots with personalized movements. A researcher commented: "You can recognize someone coming down the hall because they have a particular footfall, right? Our robot uses its limited experience to find a solution to a problem that then becomes its personalized habit, or 'personality' -- We get the dainty walker, the lazy walker, the champ... you name it." [sciencedaily.com]