Robots that understand abstract words like "soon," "many," or "behind" will learn them through their own physical experience, not from a dictionary. Most words people use are abstract—they have no single physical referent. Current robots handle concrete words like "cup" or "red" by linking them to sensors, but they fail at abstract language because they lack a body-based learning process. This project builds on developmental psychology showing that children learn abstract concepts through sensorimotor experience, not just language input. The researchers will program a humanoid robot with a cognitive architecture that transfers grounding from concrete to abstract words, using generative AI and developmental robotics methods. If successful, the work could produce robots that hold natural conversations about time, quantity, and spatial relations—not just commands about objects. This is fundamental science: it tests computational theories of how meaning arises from embodied experience. The three case studies—abstract word grounding, number word learning, and function words like prepositions—are designed to show the approach works across different types of abstraction. The project does not promise a commercial product, but deeper understanding of how machines might acquire language the way humans do, with implications for human-robot interaction in care, education, and collaborative work.
View original technical description
Language is the most natural means of communication among people, as well as for interaction between people and robots. As the great majority of words people use are abstract words, to achieve natural language interaction with robots it is crucial that machines can handle both concrete and abstract concepts. This timely project strategically builds on the recent, substantial advances in developmental psychology and embodied cognition theories on abstract concepts, on developmental robotics and AI methods for cognitive modelling, and in human-robot interaction and language use, to bootstrap our scientific and technological understanding of the grounding of abstract concepts and words in robotic agents. The project will develop and test a novel cognitive developmental architecture utilising the latest generative deep learning models and developmental robotics methods for human-robot interaction experiments on the learning of concrete and abstract words via grounding transfer. Three case studies have been strategically planned to show the breadth and robustness of the proposed approach, and to demonstrate the flexibility of the computational approach in dealing with various types of abstract concepts. The case studies focus on: (i) abstract word grounding for perceptual, motor and affective concepts, (ii) number word learning, and (iii) function words for demonstrative (deictic) prepositions, words about time and fuzzy quantifiers. This project will lead to breakthroughs in our deeper understanding of the computational approaches to the grounding and embodiment bases of language and cognition and in the design of humanoid robots which can learn to understand, and act on, conversations about a variety of abstract words, as these are intrinsically grounded in their own sensorimotor experience. The project will benefit from direct engagement with users and stakeholders of the application domains, as well as from the international advisors from both academia and industry.
Plain English summaries and category classifications on this site are generated by AI and may not perfectly reflect the original research.
Is something wrong? Let us know