ISSN 1671-3710
CN 11-4766/R
主办:中国科学院心理研究所
出版:科学出版社

• •    

社交机器人干预对儿童亲社会行为的影响: 一项三水平元分析

康丹, 胡欣怡   

  1. 湖南师范大学教育科学学院, 湖南 410081 中国
  • 收稿日期:2026-06-09 修回日期:2026-10-24 接受日期:2026-10-30

The Effects of Socially Assistive Robot Intervention on Children’s Prosocial Behavior: A three-level Meta-Analysis

Kang   

  1. , 410081, China
  • Received:2026-06-09 Revised:2026-10-24 Accepted:2026-10-30

摘要: 本研究以0—12岁儿童为对象,基于社会性互动整合视角,采用三水平元分析方法,考察社交机器人干预对儿童亲社会行为(分享、帮助、合作、安慰等)的影响效应及其边界调节机制。通过系统检索与筛选,共纳入29篇中英文实证研究,包括42个效应量(嵌套于29个独立样本),2487名被试。其中,18篇(30个效应量,1863名被试)报告亲社会行为,纳入核心分析;其余11篇(12个效应量)测量情绪、信任等亲社会行为相关构念,两部分合并(29篇,42个效应量,2487名被试),仅作探索性分析,不纳入核心结论。结果显示,社交机器人干预对儿童亲社会行为具有显著正向效应(Hedges’g = 0.73,95%CI[0.27, 1.19],p = 0.0039),但异质性较高(I2 = 93.09%),效应强度受研究内设计特征影响。调节效应分析显示,对照组类型在单因素与多因素中均显著,是相对稳健的调节变量,机器人干预相对无干预的增益较明显,相对真人互动的额外增益未获统计支持,宜定位为补充而非替代。互动方式在单因素和多因素模型中均显著,但互动方式的小组模式仅含5个效应量,估计稳定性有限;干预模式仅在单因素分析中达到显著,证据不一致,因此,这两个变量当前均不作为稳健调节变量。被试群体类型及目标指向在当前样本中未达显著水平。上述结果表明,社交机器人干预总体能促进儿童亲社会行为,效应方向较稳定,但效应强度高度依赖对照条件,并可能受互动方式与干预模式影响,不宜简单概括为普遍存在的大效应。研究结果可为智能教育场景下优化社交机器人交互设计、提升干预实践效益提供参考。

关键词: 社交机器人干预,亲社会行为,元分析,三水平模型

Abstract: This three-level meta-analysis examined the effects of socially assistive robot (SAR) interventions on prosocial behavior in children aged 0–12 years and explored the boundary conditions that moderate these effects from the perspective of social interaction theory. A systematic search and screening yielded 29 empirical studies and 2487 participants published in Chinese and English, comprising 42 effect sizes. Of these, 18 studies reported core overt prosocial behavior (30 effect sizes) and were included in the primary analysis. The remaining 11 studies measured prosocial behavior-related constructs, such as emotion and trust (12 effect sizes). The two sets of studies were combined (29 studies, 42 effect sizes, 2487 participants) only for exploratory analysis, which is reported in Appendix D and is not included in the core conclusions. Results showed that social robot interventions had a significant positive effect on children's prosocial behavior (Hedges’g = 0.73, 95%CI[0.27, 1.19], p = 0.0039), but heterogeneity was high (I² = 93.09%), indicating that effect magnitude was influenced by within-study design features. Moderator analyses showed that control condition was significant in both single-factor and multi-factor models and was therefore regarded as a relatively robust moderator: effect sizes were significantly larger under the no-intervention control condition than under the human-comparison condition; the gain relative to no intervention was relatively evident, whereas the additional gain relative to human interaction was not statistically supported, suggesting that SAR interventions should be positioned as a supplement rather than a replacement. Intervention mode reached significance only in the single-factor analysis (auxiliary vs. leading, p = 0.0197) and was not significant in the multi-factor model (p = 0.1519); the evidence was inconsistent, and this variable was therefore not regarded as a robust moderator. Interaction format was significant in both single-factor and multi-factor models, but the group-mode subgroup included only 5 effect sizes, and its estimate stability was limited; child age and self-disclosure type were not significant in either model. Therefore, interaction format and intervention mode were currently not regarded as robust moderators. Participant population type and target orientation did not reach statistical significance in the current sample. Overall, social robot interventions can promote children’s prosocial behavior, and the direction of the effect was relatively stable; however, effect magnitude depended heavily on the control condition and may be influenced by interaction format and intervention mode. The effect should not be simply generalized as a universally large effect. These findings provide a reference for optimizing social robot interaction design and improving intervention effectiveness in intelligent educational settings.

Key words: socially assistive robot intervention, prosocial behavior, meta-analysis, three-level model