It is likely that very soon robots will live in houses with humans, helping older people to live independently. But to do this, they will have to learn how to do all the little work that people could do without hesitation. Many modern artificial intelligence systems are trained to perform specific tasks by analyzing thousands of signed images of a particular action. While these methods are helping to solve more and more complex problems, they still only deal with very specific problems and require a lot of time and processing power to learn.
If a robot takes care of elderly people, the problems of such work will be very different when compared to typical situations in the learning process. During the day, robots have to do a lot of things, from making tea to changing bedding while talking. These are challenging tasks that get harder in combination. No two houses are alike, which means robots will have to quickly learn and adapt to their environment. And as is often the case, if you live with someone else, things tend to migrate. The robot will have to learn to find them on its own.
One approach is to develop a lifelong learning robot that can store knowledge based on experience and devise ways to adapt and apply it to new tasks. Once you learn how to make a cup of tea, these skills can be applied to coffee as well.
The human brain learns throughout its life, constantly adapting to complex and changing conditions and daily solving a wide range of problems. Simulating how humans learn could help design robots that we can interact with naturally, as if we were with another person.
Simulation of child development to train the robot
The first question to ask when you start modeling people is where to start? Alan Turing, a renowned mathematician and pioneer in artificial intelligence, once said:
He compared a child's brain to an empty notebook that can be filled in during education and developed an intelligent adult "system." But what should be the child's age for modeling? What knowledge and skills do you need to build first?
Newborn babies are very limited in what they can do and how they perceive the world around them. The muscle strength in the child's neck is not enough to support the head, and he has not learned to control his arms and legs.
Starting from zero month - such a step can severely limit the robot. But the child's physical limitations actually help him to focus on solving a small subclass of problems, for example, he learns to correlate his eyes with what he hears and sees. These steps in the initial stages of building a child's model build his body even before he begins to understand the complexity of the world around.
The engineers applied similar restrictions to the robot, initially blocking the movement of various joints to mimic a lack of muscle control. They also adjusted the images from the robot's camera so that it "saw" the world through the eyes of a newborn - with blur and faint peripheries. Instead of telling the robot how to move, it was allowed to figure it out on its own. The advantage here is that as the calibration changes, or as limbs are damaged, the robot will be able to adapt to these changes and continue working.
Learning by playing
Studies have shown that by applying constraints in the learning process, not only does the rate at which new knowledge and skills are acquired increase, but the accuracy of what is learned also increases.
By giving the robot control over the release of constraints - giving it control over its joints and improving its vision - it is possible for the robot to control its learning rate itself. Scientists have modeled the "baby" and the first 10 months of its growth. As the robot learned to correlate movement and the sensory information it received, it acquired the stereotypical behaviors seen in infants, such as when children spend extended periods of time staring at their hands while moving.
When a robot learns to coordinate its own body, the next important milestone it passes is that it begins to understand the world around it. Play is an important part of a child's learning. She helps him explore the environment, test different possibilities, and study the results.
At first, it can be as simple as tapping a table with a spoon or putting an object in your mouth. But then it develops into building blocks of blocks or placing objects in suitable holes. All of these actions create experiences that will further provide a foundation for skills such as finding the right key to open the lock and fine motor skills to insert the key into the keyhole and then turn it.
In the future, the use of these techniques will give robots the means to learn and adapt to the challenging environments and tasks that humans take for granted in everyday life. One day robots will be able to help the elderly, but even children in kindergarten will be able to teach them.
Ilya Khel
