Developmental Cognitive Robotics modeling for Language Learning

Student thesis: Phd

Abstract

In the past two decades, robots have become an increasingly important tool to study children's cognitive development. Many skills are learnt by children since early ages, with language acquisition being one of the most important ones, providing them the means to communicate with others, as well as supporting their personal and cognitive development. Language Learning is governed by so many factors that drawing conclusions in unconstrained scenarios can be very challenging. Cognitive modelling and robots can be useful tools in the study of language learning, focusing on specific processes influencing its acquisition, like verbal and non verbal aspects, or environmental and situational influences. In this thesis we aim to study the role of factors affecting early language learning, using a developmental robotics approach to provide computational representations of the underlying processes observed with children. We conducted three cognitive robotics experiments, in which we modelled the impact of visual attention, situational contextual diversity and social cues on early language learning. These three experiments are designed to target specifically scenarios in which the learning is based on visual position variations, on situational variations, and on the physical interaction with an instructor. The main contributions of the first two studies include the formalisation in a robotic setting of two relevant scenarios identified in the Literature (visual attention and contextual diversity), and the extension and adaptation of an existing cognitive architecture to address them. These two studies were also important to verify the suitability of the extended architecture for the types of scenarios addressed in this thesis, which is then utilised as one of the main modules in the third study. In the third and final study the contributions include the design of a novel cognitive architecture, specifically conceived to process multimodal stimuli governing language learning through the physical and verbal interaction with the instructor. The model presented, in particular, processes data collected by the robot in a real world scenario in real time. We tested and validated our models in human-robot interaction experiments involving a human instructor teaching new words to the robot playing the role of child in a real world scenario, demonstrating the effectiveness of the methods presented. Our findings are in line with those observed in the Literature, with the models presented being able to successfully describe early language learning processes. Importantly, the results do not only provide conceptual evidence of the suitability of our approach for the targeted scenarios, but also show that performance achieved allow the deployment of the models in real world scenarios. To summarise, we developed three cognitive robotics approaches for word learning, relying on data collected with a humanoid robot in an interactive scenario to train and validate our models. Our results show that the combination of cognitive skills enables the robot to successfully acquire new knowledge in a real world scenario in real time, while processing different types of stimuli with varying levels of complexity.
Date of Award6 Jan 2025
Original languageEnglish
Awarding Institution
  • The University of Manchester
SupervisorAngelo Cangelosi (Supervisor) & Amanda Banks Gatenby (Supervisor)

Keywords

  • Cognitive Architectures
  • Developmental Robotics
  • Early Language Learning

Cite this

'