Multimodality and Social Interactions in AI-Enhanced Embodied Robot-Assisted Language Learning: A Meta-Analysis

Synthesis: Wang, Zhang, and Zou (2026) meta-analyze 11 empirical studies (17 effect sizes, N = 595) of AI-enhanced embodied robot-assisted language learning (RALL) for second-language (L2) learning. A three-level random-effects model finds a positive overall effect (Hedges' g = 0.83, 95% CI [0.46, 1.21], p < .001) with high heterogeneity (I² = 84.4%). Of six moderators tested, only robot-learner interaction format reached significance (p = .049) — group-based formats showing larger effects than one-on-one — while robot morphology, modality, autonomy, social role, and duration did not. The central finding: L2 outcomes track how robots are positioned within instruction and interaction (especially group participation), not their technical sophistication.

Context

Embodied robots are increasingly used to support second-language (L2) learning, yet evidence on when and why they help remains mixed. This meta-analysis examines how robot design features and instructional conditions relate to L2 learning outcomes, drawing on Multimodal AI learning theory (MMLT) and sociocultural theory (SCT). Following PRISMA 2020 guidelines, Web of Science and Scopus were searched for empirical studies using physically embodied, AI-enhanced robots in L2 settings with sufficient statistics for effect sizes.

Key Findings

• Positive overall effect. RALL had a positive overall effect on L2 learning (Hedges' g = 0.83, 95% CI [0.46, 1.21], p < .001), but high heterogeneity (I² = 84.4%) and a small evidence base warrant caution — a provisional Benchmark, not a settled value.
• Interaction format is the key moderator. Only robot-learner interaction format reached significance (p = .049), with group-based formats showing larger effects than one-on-one interaction.
• Technical features do not drive outcomes. Robot morphology, communicative modality, autonomy, social role, and intervention duration did not significantly moderate outcomes.
• Teaching-assistant role showed the largest subgroup effect (though social role was not independently significant, and is confounded with interaction format).
• Richer multimodal/autonomous designs were not consistently linked to stronger learning.

What this means for practice

• Instructors. Judge a robot by how it is integrated into the activity, not by its hardware or AI sophistication: morphology, modality, autonomy, social role, and duration did not moderate L2 outcomes.
• Instructors. Design for group-based interaction, which was associated with larger effects than one-on-one pairings, and treat the robot as one participant in a socially organized task that supports peer Scaffolding and shared attention.
• Instructors. Invest in orchestration skill — aligning robot-mediated input with learning goals and managing turn-taking and timing — because effectiveness depended on how the robot was positioned within instruction.
• Administrators. Treat the effect estimates as provisional guidance for local procurement and adoption decisions, given the small, heterogeneous evidence base and the substantial cost of these systems.

Limitations

• Only 11 studies (17 effect sizes, N = 595) met the inclusion criteria after two studies with the lowest quality-assessment scores were removed; the authors note the evidence base is small.
• Heterogeneity was high (I² = 84.4%), so the pooled Hedges' g = 0.83 averages across dissimilar robot implementations and instructional settings.
• Only one of six moderators reached significance — robot-learner interaction format, and only marginally (p = .049) — leaving the moderator subgroups underpowered.
• Publication-bias checks (Egger's regression z = 0.97, p = .33; trim-and-fill; fail-safe N) are only approximate because several studies contributed more than one effect size and the number of studies is small.

Citation

Wang, Y., Zhang, Z. J., & Zou, D. (2026). Multimodality and Social Interactions in AI-Enhanced Embodied Robot-Assisted Language Learning: A Meta-Analysis. Educational Research Review.