Artificial Intelligence in Embodied Learning: A Systematic Literature Review
Jihye Kim, Sungwan HanABSTRACT
Background
Embodied learning (EL) is an educational approach that posits that cognition develops as learners interact with their environment through physical activity or sensory experiences. Artificial intelligence in education (AIED) supports teaching and learning for both teachers and students, for example through personalised intelligent tutoring systems. However, prior research on EL has focused primarily on students' cognitive and emotional abilities and the integration of body and task, rather than on technological elements, and research on technology‐based embodied learning (TBEL) has centred on immersive technologies such as VR and AR.
Objectives
This study aims to explore the relationship between AI technology and EL by analysing (1) the types of AI utilised in EL‐related research (algorithm or model, data modality, AI role), (2) learners' outcomes obtained from AI‐based EL, and (3) the limitations and solutions when applying AI to EL.
Method
This study is a systematic literature review based on the PRISMA 2020 guidelines, and no meta‐analysis was conducted. This study only referenced papers published since 2016 that included both AI technology and EL, were written in English, and dealt with learning processes mediated by physical activity. A study was treated as within scope only where the physical activity was tied to the concept to be learned. Papers were excluded if they did not specify the AI algorithm or model, were not peer‐reviewed, were secondary studies, were not in the field of education, were unavailable in full text, emphasised only simple skill acquisition in physical education or music, or emphasised only simple speaking activities in foreign language learning. Searches were conducted on 16 September 2025 in ACM, ERIC, and Web of Science. Given the diverse types of included studies, risk of bias was assessed using the MMAT2018 tool. Due to the varied nature of the data, a multi‐faceted data synthesis was performed. Co‐occurrences between data modality and AI role, and between AI role and learner outcomes, were visualised using heatmaps. The flow between the three variables was represented using a Sankey diagram. Finally, all frequencies were recalculated after excluding papers appraised as Low or not appraised, to test which patterns depended on weaker evidence.
Results
A total of 23 papers were reviewed. As multiple data points could be extracted per item, the relative proportion was calculated against the number of coding instances per item. The most frequently used AI algorithm/model in the EL context was deep learning, and the data modality was image data. Where multiple data types were used, it was tagged as multimodal, accounting for 34.8%. Analysis and evaluation roles were most frequently extracted, and cognitive outcomes were reported most often. Key limitations identified included restricted participant numbers due to research design and studies conducted in controlled environments. Among papers appraised as High, the most frequent pathways ran from multimodal data to recognition and detection and from sensor data to analysis and evaluation, and cognitive outcomes were the most frequently reported. When papers appraised as Low or not appraised were set aside, the concentration of AI in analytical roles persisted, whereas the evidence for AI as an interactive or generative participant weakened.
Discussion
This review describes reported patterns rather than pooled effect estimates, as no meta‐analysis was conducted. Exclusion of grey literature may introduce publication bias. Furthermore, the potential risk of selection bias exists as literature was collected based on inclusion/exclusion criteria. This study is significant in that it identified how AI is utilised within the EL context and the learner outcomes reported alongside it, particularly as the scope of AIED broadens.