Multi-Robot Collaborative Interaction in Children’s Education

Jian Huang

School of Computing and Data Science, Xiamen University Malaysia, Jalan Sunsuria, Bandar Sunsuria, 43900 Sepang,

Selangor Darul Ehsan, Malaysia

Keywords: Human-Computer Interaction, Multi-Robot Collaborative System, Children’s Education, Emotion

Recognition, Adaptive Strategy.

Abstract: With the development of artificial intelligence and robotics technology, multi-robot collaboration systems

have shown great potential in children's education. This review introduces the concepts of human-computer

interaction, multi-robot collaborative systems, and children's emotional states in the field of education. It

explains the limitations of single-modal technology and emphasizes the advantages of multimodal technology

by comparing and analysing single-modal and multimodal technologies. In addition, this review discusses the

task division, dynamic role allocation, and interaction regulation strategies of multi-robot systems in the

classroom. Finally, it points out that real-time response and child privacy protection still remain challenges

that need to be addressed urgently. It is recommended to optimize the system architecture through means such

as model pruning, introduce federated learning and build a complete privacy protection protocol to ensure the

security of children's data. This review provides theoretical and technical references for the future application

of multi-robots and multimodal perception technologies in personalized education.

1 INTRODUCTION

With the rapid advancement of technology, robotics

has emerged as a significant area of interest. In

contemporary times, robotics extends beyond the

realm of industrial automation and is increasingly

recognized for its potential in children's education. A

study conducted by Kay et al. on an international

scale revealed that during interactions with the

Cozmo robot, children not only exhibited a wide

range of playful behaviours but also regarded the

robot as a "brother" or a "pet" (Kay et al., 2023). In

another study, Yang et al. implemented the NAO

robot in classroom settings for children with autism

and observed substantial improvements in attention,

classroom communication, and positive emotional

expression (Yang et al., 2024). In conclusion, these

studies have demonstrated that the application of

robots in the field of children's education has a

promising future.

However, most educational robots today operate

as single-robot systems, limiting their ability to

provide personalized instruction. This limitation

restricts their effectiveness in delivering personalized

https://orcid.org/0009-0001-6281-6385

instruction and the capacity to execute multiple tasks.

Multi-robot cooperative systems, which leverage

coordinated behaviours across multiple robots,

efficient data transmission, and a low rate of

information errors, have emerged as a focal point in

contemporary research. The collaboration of multiple

robots can perform certain tasks that a single robot

cannot accomplish, such as dynamic role allocation

and communication collaboration. Wu and Suh

highlighted in their research that integrating learning

mechanisms into multi-robot cooperative systems

facilitates achievements such as collaborative

decision-making, task decomposition, and situational

awareness (Wu & Suh, 2024). By systematically

reviewing various robot learning methods (including

reinforcement learning, imitation learning, and

transfer learning), they concluded that robots still

possess a certain degree of universality in the current

complex educational environment.

Currently, although multi-robot collaborative

systems have demonstrated great potential in

education, they still face issues regarding the

reliability and accuracy of identifying children's

emotional states. The personalized interaction ability

Huang, J.

Multi-Robot Collaborative Interaction in Children’s Education.

DOI: 10.5220/0014355500004718

Paper published under CC license (CC BY-NC-ND 4.0)

In Proceedings of the 2nd International Conference on Engineering Management, Information Technology and Intelligence (EMITI 2025), pages 365-370

ISBN: 978-989-758-792-4

365

of educational robots largely depends on the precise

identification of children's cognitive states.

The multi-robot collaborative system, with multi-

source data sharing and multi-angle monitoring, can

intelligently adjust learning strategies based on

children's cognitive states and emotional changes,

thereby providing more flexible teaching support.

Research shows that even with simple adaptive

strategies based on rules (such as time management

and task reminders), robots can significantly enhance

children's learning effectiveness and engagement

(Rosenberg-Kima et al., 2020).

This study mainly explores how a multi-robot

collaborative system can work in real time and

achieve adaptive adjustments in an educational

environment. The research will first introduce the

core concepts in this field. Then, it will compare and

analyse the application of single-modal technology

and multimodal technology in emotion recognition.

Next, the study focuses on how robots can achieve

adaptive interaction technologies such as task

division and dynamic role allocation in the field of

education. Finally, this study identified the

limitations of current technologies and envisioned the

potential of educational robots to enhance children's

learning experience in the future. It is hoped that this

paper can provide theoretical support and technical

references for the promotion of robots in education in

the future.

2 CORE CONCEPTS

2.1 Children's Emotions and Education

The emotional state of children is a key factor in the

quality of robot-assisted education. The robot system

can make personalized adjustments to teaching based

on these identifications of children's emotional states,

thereby enhancing learning outcomes and increasing

students' acceptance. Beck et al. combined facial

expressions, gaze patterns, and deep learning models

to construct a multimodal emotion recognition system

for early education (Beck et al., 2023). The

experimental results of Laban et al. showed that

participants expressed emotions more naturally and

used richer language under the guidance of robots

with cameras and voice interaction, and their

cognitive control ability over negative emotions

significantly improved. Although the research

subjects were college students, it still has reference

value for children's education robot systems (Laban

et al., 2025).

2.2 Multi-Robot Collaboration System

In a multi-robot collaborative system, multiple robots

can work together in the same environment to

complete tasks. Compared to a single-robot system,

the multi-robot collaborative system has greater

scalability, fault tolerance, and flexibility. The robots

in this system can communicate and collaborate

through various mechanisms, such as wireless

network communication, distributed protocols,

collaborative algorithms and models, etc. The

research by Wu and Suh indicates that the multi-robot

collaborative system not only performs well in

dynamic environments such as search and rescue,

logistics, etc. but also its real-time role switching and

behaviour adjustment capabilities are applicable to

classroom scenarios (Wu & Suh, 2024). In addition,

this research points out that the application of hybrid

learning methods is the mainstream of future multi-

robot collaborative systems. By combining

reinforcement learning with imitation learning, the

system can accurately respond in real-time capturing

the learner's state and environmental changes.

2.3 Human-Computer Interaction in

Children's Education

Human-computer interaction (HCI) is an

interdisciplinary field. It aims to enhance the usability

and user-friendliness of systems by studying the

interaction between systems and users. In the field of

children's education, HCI has evolved from initial

screen projections and smart whiteboards to today's

educational robots. These robots not only facilitate

the transmission of knowledge, but also can

understand the emotions of children and adjust the

teaching pace accordingly. Yang et al.'s experiment,

found that through regular waving and LED eye

contact methods, the online attention duration of

autistic children increased by 29.22%. The rate of

children's smiling responses triggered by the robot's

praise behaviour reached 62.17%, which was 37%

higher than that in a regular classroom (Yang et al.,

2024). This immersive interactive classroom based on

emotional feedback helps maintain students'

concentration and enhance their learning interests.

3 EMOTIONAL PERCEPTION

TECHNOLOGY

Emotion perception is the core capability for

educational robots to achieve dynamic adjustment

EMITI 2025 - International Conference on Engineering Management, Information Technology and Intelligence

366

and personalized interaction. Only by establishing

real-time and precise perception capabilities can the

system achieve dynamic adjustment and thereby

enhance learning effectiveness. Currently, emotion

perception technology is mainly divided into single-

modal and multi-modal types.

3.1 Single-Modal Technology

Single-Modal technology refers to the technique that

relies solely on a single signal source for emotion

recognition. Common approaches include facial

expressions and speech.

The research by Xu et al. specifically constructed

a CNN model (Figure 1) to identify and process facial

expressions (Xu et al., 2024). The input of this model

is a 48×48 grayscale face image. The model consists

of three convolutional layers (using 32, 64, and 128

convolutional kernels respectively). After each

convolutional layer, there is a 2×2 max pooling

operation, which is used to compress the feature maps

and retain key information. Finally, through a fully

connected layer and a classifier, the image is

classified into one of the six emotions (Happy,

Confused, Bored, Claim, Excited, Anxious).

Experimental results show that the recognition

accuracy of this model for all six emotions is above

80%, with an average accuracy of 85%. This indicates

that CNN has strong capabilities for extracting and

classifying complex facial expressions, and can be

more accurate in emotion recognition for children.

Figure 1: CNN Model (Xu et al., 2024)

In order to better understand the emotional

characteristics in speech, researchers usually use Long

Short-Term Memory (LSTM) models to analyse

language. Xu et al. also constructed an LSTM-based

system (Figure 2) in the part of speech emotion

recognition (Xu et al., 2024). This model first extracts

Mel Frequency Cepstral Coefficients (MFCC) of each

speech segment as input features, then sends these data

to two LSTM layers (with 128 and 64 units

respectively), and finally classifies them into one of

the six previous emotions through a fully connected

layer and a classifier. This experiment result, found

that the recognition accuracy of each emotion by the

LSTM model is slightly higher than that of the CNN

model, with an average recognition accuracy of 87%.

This indicates that LSTM has greater robustness when

dealing with scenarios involving frequent language

communication such as education.

Figure 2: LSTM Model (Xu et al., 2024)

Single-modal models such as CNN and LSTM

achieved relatively high accuracy rates in ideal

experimental conditions for emotion recognition.

However, in real educational scenarios, the situation is

far more complex than in experiments, and their

accuracy rates may decrease in real environments.

Firstly, if only analysing from one direction and

directly drawing conclusions, it is prone to cause the

model to misjudge the emotional status of students.

For instance, some students may choose to remain

silent even if they have doubts during learning.

Secondly, external environmental factors (such as

light, noise, etc.) also affect the recognition effect.

Therefore, in order to more accurately identify users'

emotions, need to analyse them from multiple

dimensions.

3.2 Multimodal Technology

Conversely, multimodal technology benefits from its

integration of information from multiple distinct

sensory channels (such as vision, voice, brain waves,

etc.). It enhances the accuracy and robustness of

recognition, enabling the system to more

comprehensively capture the user's emotional state. It

also enables better understanding and adaptation to

different contexts and situations, thereby allowing for

more flexible adjustment of teaching strategies.

Multi-Robot Collaborative Interaction in Children’s Education

367

CNN-LSTM is a typical multimodal fusion

model. In the experiments conducted by Xu et al.,

they proposed a multimodal emotion recognition

model (Figure 3) that combines CNN and LSTM (Xu

et al., 2024). It extracts the facial image features and

acoustic features of users in accordance with the

previous path and processes them in parallel.

However, before entering the fully connected layer,

the features extracted in both directions will be

concatenated together. Finally, the emotion

classification is completed through the classification

layer. Compared with the experimental data of the

previous single-modal models, the CNN-LSTM

model has higher recognition accuracy in each

emotion category than the single-modal model. Its

average recognition accuracy even reaches 90%.

Moreover, this model can reduce the task completion

time of users by 14.3% and the error rate by 50%.

This indicates that the multimodal fusion model of

CNN-LSTM for emotion recognition is not only

technically effective but also has practical

educational application value.

Figure 3: CNN-LSTM Model (Xu et al., 2024)

Another multimodal model (Figure 4) was

developed by the Wang team, which combines

electroencephalogram (EEG) signals and facial

expressions for recognition (Wang et al., 2023). The

team first uses a pre-trained convolutional neural

network to extract image features from facial videos.

Besides, they also introduce an "attention

mechanism". This is used to highlight those moments

of facial expressions that better reflect the true

emotions. Meanwhile, another convolutional neural

network structure extracts spatial features from EEG

signals. Ultimately, these features, after being

integrated, will be sent to the classifier for emotion

determination. This integration method can better

capture the subtle changes in human complex

emotions, thereby revealing the "emotional disguise"

of children. In actual tests, the accuracy rate of

emotion classification of this model reached over

96%, significantly surpassing the traditional methods

that only use a single modality. However, due to the

high cost of the equipment required by this model,

this may limit its promotion in ordinary schools.

Figure 4: Deep Learning Model (Wang et al., 2023)

4 ADAPTIVE MULTI-ROBOT

COLLABORATION IN

CHILDREN’S EDUCATION

The multi-robot collaborative system has great

potential in enhancing children's learning experience.

Due to its ability to enable multiple robots to perform

tasks such as task allocation and role-playing, in the

face of complex and ever-changing educational

environments, the system can quickly make adaptive

adjustments. This is of great help for personalized

teaching and creating an immersive experience for

children.

EMITI 2025 - International Conference on Engineering Management, Information Technology and Intelligence

368

4.1 Task Division and Dynamic Role

Allocation

Just as Park et al. have verified in the multi-robot task

allocation, efficient task allocation and dynamic role

switching are the key factors for enhancing the

collaborative ability of the system (Park et al., 2021).

In the classroom, robots can be assigned to either

explain the course progress or promptly identify the

emotional states of the children. By transmitting the

emotional data to the teaching robot, the robot can

adjust the teaching pace according to the emotions.

The ability of dynamic role switching enables the

robots to switch tasks according to the needs of the

classroom. When students face a classroom test, the

robot will transform into a "supervisor" for

inspection. The functions of task allocation and

dynamic role switching help enhance the classroom

immersion of each child and improve their learning

experience.

4.2 Strategy Learning and Adaptive

Mechanism

The adaptive strategy refers to the mechanism by

which robots can continuously provide feedback and

adjust their own behaviours based on the real-time

emotional states of children. Through reinforcement

learning or imitation learning, robots can

continuously optimize their teaching decisions. Take

the experiment by Rosenberg-Kima et al. as an

example (Rosenberg-Kima et al., 2020). In this

experiment, the NAO robot could detect the activity

level of group members' discussions and the

completion degree of tasks to adjust the frequency of

the robot's questioning, encouragement, and time

reminders in real time. From the results, it can be seen

that these interactive behaviours have significantly

improved the quality and participation level of the

group's meeting discussions. Similarly, in the context

of children's education, robots can continuously

improve their teaching decision-making models

through the cycle of teaching and learning.

4.3 Emotional Perception and

Interaction Regulation Strategies

Accurate emotion perception is the prerequisite for

interactive regulation. Robots need to identify

children's emotions from multimodal data and then

execute corresponding strategies. For instance, when

it detects expressions of "confusion" or "anxiety", the

robot can simplify the explanation content. When

encountering "depressed" emotions, the robot can

offer encouragement or suggestions for rest. This

mechanism helps maintain children's interest in the

class and provides a more immersive learning

experience.

5 MANUSCRIPT PREPARATION

5.1 Challenge

5.1.1 Theory Challenge

There are still some challenges in the application of

multi-robot collaborative systems in the field of

children's education. The review by Dahiya et al.

summarizes that the system mainly faces

requirements for real-time performance and accuracy

in theory (Dahiya et al., 2022). The current models

make it difficult to provide feedback at the

millisecond or even microsecond level. In an

educational environment, if the delay is too large, it

may lead to a "dead silence" situation. Secondly, if

the robot's accuracy in recognizing children's

emotions or behaviours is insufficient, it may cause

incorrect strategy judgments and execute incorrect

teaching actions. This will exacerbate the children's

confusion.

5.1.2 Data Challenge

Data challenge is also a major problem that needs to

be considered in the field of children's education. The

recognition of children's emotions by robots requires

the collection of their characteristic signals. This

often faces dual challenges of privacy and ethics.

Berson et al. emphasized in their review that the

existing systems still have deficiencies in protecting

children's sensitive data (Berson et al., 2025). This

may lead to risks such as the leakage of children's

data, the commercialization of children's data, and so

on.

5.2 Future Directions

Future educational robots should mainly adopt multi-

robot collaborative systems and incorporate adaptive

strategy mechanisms. A multimodal model is used to

more accurately identify children's emotions and

issue appropriate decisions. The system can optimize

the algorithm through methods such as model pruning

to meet the low latency requirements. The adaptive

closed-loop model and continuous learning

mechanism are utilized to continuously optimize the

robot's role division and collaboration capabilities. In

Multi-Robot Collaborative Interaction in Children’s Education

369

addition, the system can also introduce federated

learning. This method trains models locally on the

school end and the home end, and only aggregates

and encrypts parameters to the central server (Hridi et

al., 2024). This can effectively prevent the leakage of

children's data and improve the fairness of the model.

6 CONCLUSIONS

This paper systematically reviews the potential of

multi-robot collaboration systems in the field of

children's education. Multi-robot collaboration

systems can utilize advanced emotion recognition

technology and adaptive interaction strategies to

enhance flexibility and personalization in the

classroom environment. This article first introduces

the relevant concepts, emphasizes that the multi-robot

collaborative system has a promising future in the

field of children's education, and explains the

limitations of the single-robot system.

Then, this paper compares and analyses single-

modal and multimodal recognition technologies.

Although CNN and LSTM can achieve relatively

accurate recognition in experimental environments,

they are still prone to be affected by noise and

complex emotional states in real-world environments.

The multimodal models such as the CNN-LSTM

model and the EEG-facial fusion model combine

visual, auditory and neural signals, which can achieve

higher recognition accuracy.

After that, this paper explores the application of

adaptive strategies in the classroom environment. In

the future, robots can continuously optimize the

classroom rhythm and adjust their interaction

behaviours with children based on an adaptive closed-

loop model.

Finally, this paper summarizes some of the

remaining challenges in this field. For instance, the

demand for high real-time performance and accuracy

in the classroom environment, the risk of sensitive

data leakage of children, and ethical security. Future

research should focus on model optimization to build

a system architecture capable of achieving

millisecond-level response. Additionally, encryption

mechanisms such as federated learning should be

introduced and a comprehensive privacy protection

protocol should be established to ensure the data

security of children.

In conclusion, this field still needs to conduct

empirical research in real classroom settings to collect

more data. Only on the basis of interdisciplinary

collaboration and regulatory guarantees, the multi-

robot collaborative system can truly achieve

sustainable promotion in the education field and

provide personalized and effective learning support

for more children.

REFERENCES

Beck, E., Bockelmann, C., & Dekorsy, A. (2023). Semantic

information recovery in wireless networks. Sensors,

23(14), Article 6347.

Berson, I. R., Berson, M. J., & Luo, W. (2025, March).

Innovating responsibly: Ethical considerations for AI in

early childhood education. AI Brain Child, 1(1), 2.

Dahiya, A., Aroyo, A. M., Dautenhahn, K., & Smith, S. L.

(2023). A survey of multi-agent human–robot

interaction systems. Robotics and Autonomous

Systems, 161, 104335.

Hridi, A. P., Sahay, R., Hosseinalipour, S., & Akram, B.

(2024, May). Revolutionizing AI‑assisted education

with federated learning: A pathway to distributed,

privacy‑preserving, and debiased learning ecosystems.

AAAI Student Symposium, 3(1), 297–303.

Kay, L., Brandsen, S., Jacques, C., Stocco, F., & Zaffaroni,

L. G. (2023). Children’s digital and non‑digital play

practices with Cozmo, the toy robot. M/C Journal,

26(2).

Laban, G., Wang, J., & Gunes, H. (2025). A robot-led

intervention for emotion regulation: From expression to

reappraisal. arXiv preprint arXiv:2503.18243.

Park, B., Kang, C., & Choi, J. (2021). Cooperative

multi‑robot task allocation with reinforcement learning.

Applied Sciences, 12(1), Article 272.

Rosenberg‑Kima, R. B., Koren, Y., & Gordon, G. (2020).

Robot‑supported collaborative learning (RSCL): Social

robots as teaching assistants for higher education small

group facilitation. Frontiers in Robotics and AI, 6, 148.

Wang, S., Qu, J., Zhang, Y., & Zhang, Y. (2023).

Multimodal emotion recognition from EEG signals and

facial expressions. IEEE Access, 11, 33061–33068.

Wu, B., & Suh, C. S. (2024). State-of-the-art in robot

learning for multi-robot collaboration: A

comprehensive survey. arXiv preprint

arXiv:2408.11822.

Xu, Y., Lin, Y.-S., Zhou, X., & Shan, X. (2024, June).

Utilizing emotion recognition technology to enhance

user experience in real‑time. Computers and Artificial

Intelligence, 2(1), 1388.

Yang, Q., Lu, H., Liang, D., Gong, S., & Feng, H. (2024).

Surprising performances of students with autism in

classroom with NAO robot. arXiv preprint

arXiv:2407.12014.

EMITI 2025 - International Conference on Engineering Management, Information Technology and Intelligence

370