Human babies are unique in the animal world in that they are born with no ability to care for themselves or walk. Other animal species have offspring that are ready to run with the herd within a very short time of being born. Animals like giraffe or wildebeests need no instruction to walk; they just do it. Researchers at the USC Viterbi School of Engineering have created an AI-controlled robotic limb that requires no programming to learn to perform tasks.
The leg, which uses animal-like tendons, can be tripped and recover from that trip before the next footfall despite that it was never programmed to be able to do so. The bio-inspired algorithm can learn to walk by itself after only 5 minutes of unstructured play, then adapt to other tasks with no additional programming.
The team thinks that this algorithm might allow the creation of prosthetics and robots that can interact with complex and changing environments, such as found in space exploration and search-and-rescue. The researchers note that it usually takes the equivalent of months or years of training for a robot to be ready to interact with the world.
The approach the team uses first allows the robot to understand the environment in a free play process. The random movements of the robot leg in the experiment allowed the robot to build an internal map of the limb and its interactions with the environment. The robots the team makes can learn-by-doing without prior or parallel computer simulations to guide its learning.
The robots learn a solution that works well enough for them allowing them to be more flexible and learn faster. This produces robots with personalized movements that are recognizable by their gait. The team sees its new algorithm having potential uses in assistive devices such as robotic limbs and exoskeletons.