Scientists from the Faculty of Electrical Engineering at the Czech Technical University in Prague are trying to understand one of the key questions in the development of human cognition: how young children form their perception of the body and the relationship between cause and effect. With the help of the humanoid robot iCub, they are now simulating a famous experiment by American psychologist Carolyn Rovee-Collier, in which six-month-old children learn that the movement of their own limbs can influence the world around them.
The team led by Associate Professor Matěj Hoffmann placed the iCub robot in a crib equipped with a simple toy that moves when the robot pulls a string with one of its limbs. Using an artificial neural network, iCub learns to recognize which of its limbs is moving the object and begins to use it more actively—just like a human child. The result is an algorithm that mimics early sensory-motor development.
“We are interested in what an artificial brain must contain in order to behave similarly to a child's brain,” explains Matěj Hoffmann, who heads the humanoid robotics laboratory at the Department of Cybernetics at CTU. “The goal is not only to develop robots, but also to gain a deeper understanding of the development of human cognition.”
Psychologists who appreciate the contribution of robotics to modeling child development are also involved in the research. “The experiment with the iCub robot gives us a unique opportunity to test the theoretical assumptions of developmental psychology on a completely controllable system,” says Dr. Sergiu T. Popescu, a developmental psychologist working in the research group. “Thanks to the robotic model, we can better understand how a child learns to distinguish that its movements affect the world. “Agency” is a crucial moment in development when a child first realizes that it can change its environment through its own actions.”
Significance of the research: modeling cognition from the ground up
First results from the research by the FEE CTU team come from a simulated environment where the child was simplified to a direct connection of artificial “motor” neurons with movements of virtual arms and legs (preprint available at A computational model of infant sensorimotor exploration in the mobile paradigm) This work presents the first detailed computational model that mimics a key mechanism of infant cognition. The simulation of the mobile paradigm shows that the robot can learn to move the specific limb that causes a response, similar to a human infant. Current experiments are working with a full-fledged baby robot in a crib. The robot's body, position on the back and the whole context are crucial for understanding how infants learn sensorimotor skills.
The model uses a neural network with a prediction mechanism and an exploration module. The robot actively seeks out movements whose results are surprising to it, thereby learning more about its body and environment. In a series of “ablation studies,” the authors show that without prediction, motor noise, or a sufficient number of “muscle commands,” the robot's behavior differs fundamentally from that of a human child. The model also mimics the less common phenomenon of extinction burst—a short-term increase in activity after the removal of a stimulus.
“Our results show that not only the ability to move is crucial for learning, but also the ability to make predictions and respond to their failure,” adds Matěj Hoffmann.
The model provides evidence that the basic mechanisms of learning—surprise, motivation to discover, and prediction—can arise without external rewards, purely from internal organization. The authors plan to further develop the model to include factors such as boredom, attention, and multiple types of stimuli. This approach opens up new possibilities not only for the development of autonomous robots, but also for research into early cognitive development in children.
What international experts say
“The so-called "mobile paradigm" is a simple experiment used by developmental psychologists to study how babies use feedback from the environment to learn how to move their limbs and explore the world. A "mobile" is suspended above the infant's crib and attached to the infant's foot, so that when the infant kicks, the mobile moves. Over time, babies learn to kick more to see the mobile move, showing they remember and understand the connection. Testing a humanoid robot in this paradigm is a useful way to evaluate the algorithms used by robots to move their own bodies and explore their environment. The aim is to make robots learn in the same highly adaptable way that human infants do,” says J. Kevin O'Regan, Emeritus Director of Research from Integrative Neuroscience and Cognition Center, Université Paris Cité
iCub robot: a child among robots
ICub is a humanoid robot designed by the Italian Institute of Technology (IIT) for research into cognition and interaction with the environment. It measures just over one meter, is the size of a four-year-old child, and its body is controlled by 53 electric motors. It perceives the world through two cameras (eyes), microphones (hearing), and thousands of touch sensors built into its “electronic skin.” It is this combination that makes it an ideal candidate for research that combines robotics, artificial intelligence, and developmental psychology.
A group of researchers from the Department of Cybernetics at the Faculty of Electrical Engineering of the Czech Technical University in Prague acquired the iCub robot as part of a project of the Research Center for Informatics. It is the first robot of its kind in the Czech Republic and one of the few that runs on open software, which allows for its further development in cooperation with the international community.
A video recording of the experiment with the robot is available here.
