Meta-analysis shows moderate positive learning effects from physically embodied educational robots in classrooms

Physically embodied educational robots (PERs) are increasingly integrated into classrooms, yet evidence on their impact on children's learning remains heterogeneous. This meta-analysis synthesizes 34 independent studies (N = 3665) to estimate the overall association between PER-supported instruction and learning outcomes and to examine boundary conditions interpreted through a Cognitive Load Theory (CLT) lens. Overall, PER-supported instruction was associated with a moderate positive effect on learning outcomes (r = 0.31, 95% CI [0.24, 0.38], p < .001). Moderator analyses indicated larger effects when robots were implemented in active instructional roles (e.g., collaborative or teaching-by-learning), when programmable educational kits were used, when interventions were delivered as longer-duration programs (> 4 weeks), and when PERs were compared with traditional materials/tools rather than human instructors or other comparison conditions. In contrast, we did not detect reliable moderation by educational stage, country, session length, or outcome domain in this corpus. Interpreted through CLT as an organizing lens, the moderator pattern is broadly consistent with the possibility that PERs may support learning when they help structure task processing and reduce avoidable cognitive demands in well-scaffolded activities. Practically, the findings suggest that more robust gains are most likely when robots are embedded in sustained, structured learning designs rather than brief demonstration-only lessons.