Study shows robots learn better by mimicking early learning experiences of infants
The research opens new avenues for understanding how early experiences shape cognitive and motor development, with potential implications for both robotics and AI.
A collaborative study from Birkbeck, New York University, and the University of Texas, Austin, has found a link between infants' everyday variable experiences to their early learning of foundational skills like walking or talking. Specifically, the study used robot simulations to understand how babies learn to walk from their real-world experiences.
The research team used robots programmed to learn from real infant movement patterns. By testing the robots in simulated environments, including engaging them in football games, also known as RoboCup, the researchers assessed the impact of two things. First, the variability of walking paths – in which robots learned different walking paths to see how it affected their ability to move overall. Secondly, the subjective consequences for errors on learning outcomes – in which robots were either penalised or not penalised for any mistakes they made in their learning.
The study found that robots trained on highly variable, real infant walking paths (recorded in the lab for 20 sessions with infants running around the room) led to more wins in robot football games, than robots on traditional paths (e.g., straight lines like how physical therapists train babies on a treadmill when they have delays in walking). Similarly, robots trained with a minimal penalty for falling won more football matches than robots trained with a high penalty for falling.
Dr Ori Ossmy, Reader in Developmental Cognitive Neuroscience at Birkbeck commented:
“A ‘penalty for error’ means facing a consequence or losing something whenever a mistake is made, which can also refer to how much babies care about falling when they are first learning to walk. When babies are learning to walk and show little care about falling, it helps them learn and develop crucial skills. Likewise, if a person is learning a new language and doesn't mind getting things wrong, they will likely learn better. The same is true for sports or any learning in life.
These findings underscore the importance of diversity and low consequences for errors in fostering adaptive behaviour. They suggest that infants' spontaneous, varied activities combined with their low penalty for error, plays a critical role in laying the groundwork for essential skills like walking, talking, and social interactions."
Beyond the implications for infant development, the study highlights the potential of robotics in elucidating complex learning processes. Dr Ossmy envisions a reciprocal relationship between robotics and developmental psychology, where insights from one field inform the other.
"Robots offer a unique perspective on development, mirroring the challenges faced by infants in coping with a body embedded in an ever-changing environment," Dr Ossmy added. "Similarly, observations of infant behaviour may inspire advancements in artificial intelligence, particularly in achieving functional, adaptive performance."