International Journal of Reconfigurable and Embedded Systems (IJRES)

Vol. 14, No. 1, March 2025, pp. 35~47

ISSN: 2089-4864, DOI: 10.11591/ijres.v14.i1.pp35-47  35

Artificial intelligence driven robotic control system for

personalized elderly care and foot massage

Shripad Bhatlawande, Swati Shilaskar, Soham Akotkar , Anish Joshi, Zayd Ansari
Department of Electronics and Telecommunication Engineering, Faculty of Engineering, Vishwakarma Institute of Technology, Pune,
India

Article Info ABSTRACT

Article history: This research presents an electronic system for providing foot massage to the
elderly, along with artificial intelligence (AI) driven voice-controlled
Received Feb 27, 2024 conversation bot. The problem under study focuses on the elderly age group
Revised Jul 10, 2024 suffering from foot related ailments, most commonly foot pain. Also, the
Accepted Aug 12, 2024 risk of depression or anxiety is high for this age group due to social
isolation. These problems are addressed by the system under discussion
integrated with a voice assistant to converse with the elderly. The AI assisted
Keywords: conversation bot enables the elderly to make customized reminders for their
timely medications and provides general updates on essential topics. The
Artificial intelligence assistant system extends to provide the elderly, foot, and calf massage controlled with
Conversation bot mobile application. It consists of a low power motor arrangement along with
Depression a high computing system. The electronic system was subjected to trials on
Elderly care system elderly for verification and validation of the system to assess its ability of
Foot massage providing users with appropriate assistance. The trials were conducted on
twenty elderly, aged sixty, and above, living self-sufficiently with foot
related ailment. All elderly were subjected to the conversation bot along with
the foot and calves’ massage, providing subjective feedback on the system's
ability to enhance their quality of life. The subjective feedback after
quantification have demonstrated the ability of the system in improving their
living standards.
This is an open access article under the CC BY-SA license.

Corresponding Author:
Swati Shilaskar
Department of Electronics and Telecommunication Engineering, Faculty of Engineering
Vishwakarma Institute of Technology
VIT, Upper Indira Nagar, Bibwewadi, Pune, Maharashtra 411037, India
Email: [email protected]

1. INTRODUCTION
The number of elderly staying alone is increasing with an adverse impact on their health and
psychological behavior. Recent reports provide an insight regarding elderly population (aged 60 and above),
which is currently at 10.5% of India’s total population [1]. About 5.7% of the elderly population live alone
without the support of any family members [2]. This situation presents a significant problem, the absence of
companions and the unavailability of individuals to look after the elderly's health. Prominent reasons for this
unavailability might include a lack of family support or situations where the elderly are staying alone.
Additionally, hiring a trained career to provide necessary services, such as massage therapy, can be
prohibitively expensive. Addressing these challenges is crucial for improving the well-being and life
satisfaction of the elderly population. This study aims to tackle these issues by developing innovative
solutions that ensure both physical and mental assistance for the elderly, ultimately enhancing their quality of

Journal homepage: http://ijres.iaescore.com

36  ISSN: 2089-4864

life. The field of elderly care and assistance has witnessed significant advancements in recent years, primarily
driven by the integration of cutting-edge technologies like artificial intelligence (AI) and robotics.
The researchers have developed conversational robots for companionship to elderly care with
assessing the understanding of elderly and creating a dataset to check prefrontal cortex (PFC) and working
memory (WM) [3]. One of the key elements of robotic elderly care is the accuracy with which actions are
identified by the robot using data from sensors [4]. By 2050, the elderly will make up around 16% of the
global population. To provide better aged care and increase life expectancy, it is essential to monitor older
people's physical activity, particularly for fall detection and prevention [5]. For the elderly who are living
alone at home, homecare research is particularly important in solving the aging issue. Approaches like
camera-based monitoring and wearable based smart devices were tried but did not get the results as expected
[6]. The caregivers are now being trained using AI robots which use visual based feedback such as pain level
which gives trainees more experience [7]. Robotics and AI together have a bright future in elderly care and
assistive technologies.
The proposed system aims to enhance the quality of life by providing digital companionship and
massage therapy to the elderly. These objectives are achieved by integrating an electronic system of end
effectors with an AI driven conversational assistant. The assistant is designed to generate a stipulated
response by analyzing the intent and emotion behind the user’s input. A novel intent recognition system
using few-shot learning and a fine-tuned emotion classifier is proposed. The AI assistant decides the response
by classifying user input into certain predefined intent classes. Furthermore, emotional analysis of natural
language in text is performed to simulate human-like conversations with the elderly. The complex massage
system seamlessly collaborates with the AI conversational robot thereby ensuring physical and mental
assistance for the elderly. The development of the AI assistant contributes to the existing work in terms of
providing companionship to the elderly.

2. LITERATURE REVIEW
Researchers have worked on elderly care to provide an assistive framework for elderly as well as
some kind of physical aid. Further, the work done by researchers is discussed along with its outcomes. Smart
elderly care robots equipped with AI features provide personalized care, companionship creating
personalized datasets to assess PFC and WM to check current understanding of storytelling and memory and
difficulty in understanding the story, the model has the accuracy score of 89% [3]. Deep learning approaches,
such as convolutional neural networks (CNN) and recurrent neural networks (RNN), have improved the
accuracy (92.31%) of activity recognition by capturing spatial and temporal relationships, but they require
large amounts of training data and have slow training speeds [4]. The combination of fall detection and
human pose estimation using various pipelines such as MovNet, transformer, long short-term memory
(LSTM), and AlphaPose. The poorest results were obtained by MovNet 69%, while the greatest results were
attained by AlphaPose 89.22% for pose detection and transformer 90% for action categorization [5]. The
three phases of the behavior identification algorithm (BIA) are sensor event based grouping (SEG),
histogram similarity based grouping (HSG), and behavior based matching (BM). The BIA has the F1-score
of around 95% [6]. The introduction of the care training assistant robot revolutionizes caregiver training,
offering visual-based feedback in elderly care education environments. This immersive approach allows
trainees to acquire and practice essential caregiving skills, elevating the standard of elderly care services with
pain levels from 0 to 1.99 which gives the caregiver more exposure for practical applications [7].
A smart elderly assistance system which uses computer vision algorithms to identify fall condition
has been developed. The limitation of the system is the requirement of proper lighting conditions. The author
claims accuracy of 98:5%, sensitivity of 97% and the specificity of 100% for the algorithm [8]. A smart
robotic arm Emma is designed for medical massage with feedback of force applied through end effectors. It
further has a camera for exact placement of finger and palm massage manipulator to conduct feedback
control with respect to force [9]. HomeMate is a system which tries to establish ambient assisted living
around the elderly. Social robot is another assistive framework that has human robot interaction (HRI) and
tries to establish communication with elderly [10]. The author has provided an overview on 12 elderly care
systems focused on assistive care that can be provided for exercise and rehabilitation through cognitive based
assistance bot (CAs) and personal robot based assistant bot (PRAs). PHAROS is a combination of robotic as
well as event planning for elderly monitoring, while they perform physical tasks. Stevie is another assistive
robot which engages in social interactions with elderly through gestures and speech [11]. The author
discusses smart homes in the context of ‘aging of elderly in place’. A system “CareBot-assisted smart home
(CBASH)” is by the author which consists of a sensor-based network, a service processing center and a care
responsive Bot. The CareBot is equipped with LIDAR and momentary feedback for navigation and video call
functionality along with an AI assisted chat system for interaction with elderly. An improved genetic based

Int J Reconfigurable & Embedded Syst, Vol. 14, No. 1, March 2025: 35-47
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  37

algorithm (IGU) consists of various operators (fitness score (450), selection operator, and threshold (1,000))
for task planning. The author confirms that the experiment results show that the CBASH system can perceive
complicated tasks and plan for homecare assistance tasks with a quick response in case of emergencies [12].
A robotic massage positioning system which uses prediction modeling of feet acupoints guided by traditional
Chinese acupuncture empirical knowledge. A cerebellar model arithmetic computer (CMAC) neural network
to further optimize detection of acupoints with reinforcement learning and predictive models resulted in a
highly accurate acupoint coordinate system [13]. The psycho-physiological and muscular parameters were
studied.
The study concluded that massage stimulation on pressure receptors resulted in improved vagal
activity and cortisol levels [14]. The result of the study of assistive and nursing robots concluded that very
few technological assistive elderly care solutions were in use due to high costs, legal and safety issues yet to
be resolved [15]. Most of the robots were mitigation-based and emergency centric assistants [16]. A detailed
report of nursing robots with case studies to analyze the performance of robots in healthcare and nursing
environments. It also talks about the social acceptability of robotic healthcare systems within elderly circles.
The authors cite a report wherein 51% of European citizens were uncomfortable using care robots [17]. A
cost-effective radiation monitoring system is discussed for nuclear sites utilizing wireless sensor networks
(WSNs) and ZigBee technology. The system demonstrated 25% reduction in cost compared to available
commercial solutions [18]. The literature describes design and implementation of a cost-efficient smart
vehicle for environmental parameters analysis in industries, for the well-being of employees [19]. Robots
tailored for frail elderly care and their nursing modes are found to be expected to give interpersonal
interaction experience to the elderly [20]. An adaptive and aware dialogue system for elderly using virtual
assistant is developed and the need for interactive communication for the elderly is emphasized in the work
[21]. A compression wave pattern by varying frequency, amplitude and phase generated for massage [22].
Acupoint massage system was developed [23]. Physiological parameters measurement during massage on
static and dynamic condition of foot movement was implemented [24].
Comparative study of difference in the skin conductance , as a measure of relaxation was carried out
involving a physical therapist and intervention by a robot [25]. It was concluded that the manual massage
intervention outperformed robotic massage. The field of assistive technology, elderly healthcare using
robotics has seen many advancements in recent years. According to multiple research papers reviewed, some
conclusions regarding the research gap could be drawn with respect to the new advancement considered in
our system. Many systems as mentioned in the literature survey consisted of features such as - face
recognition/elder motion analysis, self-navigation using mapping and localization. Chatting and conversing
with elders, remote connectivity through smartphone, physical aid like massaging, physiological and
psychological therapy, cognitive training and exercises for elders, determining relation between perception
and action, generating an action based on object detection and graph neural network (GNN), human activity
and health monitoring. The reviewed papers implement some of the above-mentioned ideas. Many research
papers that are reviewed present mobile robots with AI technology, that are able to detect faces, used for
increasing cognitive as well as brain agility, finding relationship between perception and action but none of
them provides any kind of physical aid to the elderly people. The system proposed in this work integrates
conversational AI assistance for elderly as well as an embedded system for foot and calf massage. The
system under study provides physical aid in the form of foot and calves massage assisted by a personalized
AI assistant which will control the massage system through the mobile application and provide
companionship through conversation.

3. METHOD
The system consists of an AI-based conversational bot (robot) and a control system for foot sole and
calf massage. The entire system is connected to a mobile application via Bluetooth for user interaction using
voice and text commands. The system comprises four major components: a platform for foot which consists
of vibration motors at specific acupoints based on the foot reflexology, two medical cuffs are fixed with the
air pump for the calf massage, which inflates and deflates to generate the massage, both the systems are
connected with a fast-computing system which is controlled by a mobile application through Bluetooth as
represented in Figure 1. This system is composed of two main subsystems: conversation robot with foot sole
and calf massaging system along with a mobile application for controlling the intensity of the massage. The
Raspberry-Pi acts as the central processing unit for the system.

3.1. Development of conversation robot

The development of the AI assistant is a pivotal element of this research article. The AI is designed
to provide companionship and assistance with medication management through reminders. AI learns from
user interactions, adapting its responses and services over time. The AI translates the user's health objectives,
Artificial intelligence driven robotic control system for personalized elderly care … (Shripad Bhatlawande)
38  ISSN: 2089-4864

expressed through natural language input, into specific vibration patterns and intensities for the massage. The
conversational bot performs mainly three tasks - emotion recognition and response, processing natural
language input for massage control and setting medication reminders for the elderly. This research work
proposes a custom-built intent recognition system that is able to distinguish natural language input into the
above designated intents.

Figure 1. AI assistant driven massage control system

3.1.1. Custom intent recognition system using few shots learning

A diverse dataset is curated with multiple examples necessary for few shot learning with examples
of different phrasings, emotions, and contextual details along with their labeled intent. This data undergoes
the standard 80:20 train test split. The conversational robot is triggered by receiving a natural language input
from the user, typically in the form of a conversational query or command. Input received through voice is
converted into text using speech to text libraries like pyaudio. The input text is tokenized into words or sub
words, and contextualized embeddings are extracted using the pre-trained GPT-3 model. These embeddings
capture the semantic meaning and context of each word within the sentence. The algorithm utilizes few-shot
learning, providing the GPT-3 model with prompts that include examples of each intent. Sample input for
few shot learning:
− Intent 1: ("Converse with elderly"): "Converse with an elderly person about their day."
− Intent 2: ("Manipulate massage control"): "Adjust the robotic massage control system to a gentle
intensity."
− Intent 3: ("Set medication reminders"): "Create a medication reminder for a daily dosage of [Medication
Name]."
The GPT model is trained on these few-shot examples to adapt to the specific intent recognition
task. The model while training and fine-tuning adjusts its weights to better recognize patterns associated with
each intent. The classified intent is then mapped to pre-defined functions or tasks.

3.1.2. Emotion recognition

The algorithm for emotion classification is initiated when the intent classifier recognizes that the
user wants to have a conversation. In this study, we employed a bidirectional encoder representations from
transformers (BERT)-based emotion classification model for fine-tuning on the "dair-ai/emotion" dataset.
The dataset is split into 80% training, 10% validation and 10% testing. It consists of 6 output classes that are
(0: sadness, 1: joy, 2: love, 3: fear, 4: anger, 5: surprise) that represent six basic emotions. The dataset was
loaded using the hugging face datasets library, and tokenization was performed using the BERT tokenizer.
To ensure compatibility with the model, the maximum sequence length was determined based on the training
dataset. The dataset was then preprocessed to include padding and truncation with a specified maximum
length. The base model “BERT-base-uncased” was fine-tuned on the training data for 5 epochs. For a sample
user input of “I miss the times when their laughter echoed through these walls. Now, it's just silence, and I

Int J Reconfigurable & Embedded Syst, Vol. 14, No. 1, March 2025: 35-47
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  39

can't help but feel a bit forgotten.” The fine-tuned BERT model performs emotion classification by
generating probability distributions in the 6 output classes or emotions. The confidence scores provided by
the fine-tuned BERT model are displayed in Table 1.

Table 1. Confidence scores

Emotion Sadness (0) Joy (1) Love (2) Fear (3) Anger (4) Surprise (5)
Confidence score 0.984 0.002 0.004 0.005 0.003 0.002

3.1.3. Response generation

The emotion with the highest confidence score was employed to generate an emotionally aware
response from the conversational bot. The OpenAI GPT-3 model was deployed for response generation. The
recognized emotion, obtained from the fine-tuned BERT model, served as a prompt for the GPT-3 model.
The GPT-3 API was utilized to generate contextually relevant responses. Parameters such as max_tokens and
temperature were adjusted to control response length and creativity. The integration of the emotion
recognition and response generation models provides a comprehensive system for understanding user input
and generating appropriate and contextually relevant responses. The response generated based on the
previous example would be – “I'm sorry to hear that you're feeling this way. It's natural to miss the joyous
moments and the laughter that once filled the space around you. Remember that emotions, including
nostalgia and a sense of being forgotten, are valid. If you need someone to talk to or share your feelings with,
I'm here for you.”

3.1.4. Manipulation of robotic massage control system

Upon classifying the user's intent as "Manipulate massage control," the algorithm commences with
user input processing, employing regular expressions (regex) to identify massage-related instructions,
encompassing aspects like intensity, areas, or feedback. Extracted information is then harnessed to
dynamically adjust the parameters of the robotic massage control system. For instance, the algorithm
responds to user preferences by increasing or decreasing intensity, directs the massage system to focus on
specific areas of the body, and accommodates alterations in massage modes or patterns. This approach
ensures a responsive and tailored experience, allowing the elderly care chatbot to intuitively interpret user
instructions and execute precise adjustments to the massage system. The user_input "increase the intensity
and focus on the tibial foot area" is processed with given (1):

𝑚 = ′\𝑏( 𝑖| 𝑎| 𝑓)\𝑏′ (1)

where \𝑏 is a word boundary to ensure whole words are matched, 𝑚 is massage parameter, 𝑖 is intensity, 𝑎 is
area, and 𝑓 is feedback. (intensity|areas|feedback) is a group that matches any of the specified words
(intensity, areas, or feedback).
The regex would identify "intensity" and "areas" as massage-related instructions, and the extracted
information would be "intensity, areas". This information can then be utilized to adjust the parameters of the
robotic massage control system accordingly. Medication reminder: upon identifying the user's intent as "Set
medication reminders," the algorithm employs regex to meticulously extract pertinent medication-related
details, such as medication names, dosages, and frequencies, from the user's input. The user_input = "Set a
reminder for Azithrol at 9 AM every day" undergoes parameter extraction using (2):

𝑟 = ′\𝑏(𝑚𝑒𝑑|𝑑|𝑓𝑟𝑒𝑞)\𝑏′ (2)

where, 𝑟 is reminder parameter, 𝑚𝑒𝑑 is medication, 𝑑 is dosage, and 𝑓𝑟𝑒𝑞 is frequency.

The regex in (2) will extract medicine name, time and frequency. The extracted information is
utilized by the algorithm to dynamically adjust the parameters for setting up a medication reminder.
Subsequently, the system utilizes Bluetooth communication to transmit the medication reminder information
to a connected mobile application. The implementation involves establishing a Bluetooth connection and
sending relevant data to the mobile app, thus ensuring seamless integration between the elderly care chatbot
and the mobile platform. The algorithm concludes by generating a confirmation response to notify the user of
the successful setup of the medication reminder. The summary of the end to end execution of user input by
the AI assistant can be described in Figure 2. This process flow is crucial for the efficient functioning of the
control system and ensures that a correct decision is made for each user input. This algorithm runs in a loop
wherein each user inputs the request met with a corresponding output. The end nodes of the process as given
in the flow diagram in Figure 2, represent the corresponding outputs for every type of user input.
Artificial intelligence driven robotic control system for personalized elderly care … (Shripad Bhatlawande)
40  ISSN: 2089-4864

Figure 2. AI assistant: intent recognition and action execution

3.2. Design of foot and calf massager

3.2.1. Calf massage
The system consists of a pneumatic network for connecting air pumps with two medical cuffs. It is
powered using a lithium polymer battery pack with Raspberry Pi acting as its processing unit. The algorithm
executed on the processing unit decides the duty cycle for inflation and deflation of the medical cuffs based
on the inputs provided by the app.

3.2.2. Foot sole massage

The system has acupuncture points allocated at 7 different positions on the foot sole according to the
foot reflexology as shown in Table 2. The different acupuncture points base is connected with a series of DC
motors attached, which are responsible for enhancing the functionality of particular organs. The foot sole is
divided into primarily three sections- lateral plantar, median plantar and tibial plantar as shown in Figure 3.
The placement of vibration motors is provided in Table 2 and is performed using foot reflexology technique.
It gives information about the placement of vibration motors and the organs acupoints that are being targeted
by the vibration motors. The motors are situated with five (vib1 to vib5) in the median plantar area, one in the
tibial plantar area (vib6), and one in the lateral plantar area (vib7). The default duration of foot sole massage
is 3 minutes. The user can alter the pulse width modulation (PWM) of each of the three foot-sole sections
separately by providing voice feedback. The maximum and minimum RPMs of the vibration motors mounted
on the foot massage platform are 13,500 and 8,500, respectively.

Table 2. Vibration motor placements respective to acupoints

Vib1 Vib2 Vib3 Vib4 Vib5 Vib6 Vib7
Heart Lungs Liver Pituitary Stomach Small intestine Arm

Int J Reconfigurable & Embedded Syst, Vol. 14, No. 1, March 2025: 35-47
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  41

Figure 3. Representation of divided foot sole along with sensor positioning

Few of the most common health challenges are related to the heart, lungs, liver, pituitary, stomach,
small intestine and arms. The vibration motors are placed on the acupoints related to the above-mentioned
areas based on the foot reflexology, these points are associated with different body organs and functions. A
predefined mapping of acupoints on the feet as shown in Figure 3 is employed. The location of these
acupoints is correlated with the positioning of the vibrational motors on the massage subsystem.

𝐹𝑜𝑟𝑐𝑒 𝑎𝑝𝑝𝑙𝑖𝑒𝑑 = 𝑓(𝑥, 𝑦, 𝑧) (3)

Where: 𝑥 is user's health objective, 𝑦 is acupoint position and 𝑧 is motor capability.

The (3) describes the application of force as a function of users’ health objective, acupoint position
and motor capability connected.

𝑡 = 𝑉𝑐𝑢𝑓𝑓/𝑄 (4)

where: t is the time in minutes, 𝑉𝑐𝑢𝑓𝑓 is the capacity of each medical cuff bag in milliliters (200 ml) (two of
such cuffs used), and 𝑄 is the flow rate of the air pump in milliliters per minute (12,500 ml/min, as 1
liter=1,000 milliliters).

𝑡 = 200 ∗ 2/12,500

𝑡 = 0.032 𝑚𝑖𝑛 ~ 1.90 𝑠𝑒𝑐𝑜𝑛𝑑𝑠

By the calculations given in (4), both the cuffs get inflated in 1.9 seconds considering the 100% duty
cycle of the air pump motor. Furthermore, the motor has the capability to exert a minimum of 200 mmHg and
a maximum of 325 mmHg pressure on each of the connected medical cuffs. The default duration of calf
massage is 3 cycles of 5 seconds each, considering the instance when massage is requested through mobile
application. After completion of massage, the user may request for another cycle of massage if necessary.
The user can alter the intensity of inflation and deflation calf cuffs through mobile application. The intensity
levels are classified into three states: low, medium, and high. When the user clicks the increase/decrease
intensity button, the status of the massage changes in relation to its present state. The algorithm provides
control and response generation for the massage device via a mobile application and voice input. It handles
commands for initiating foot sole massage, adjusting the intensity of calves’ massage, and responding to user
prompts as mentioned in Algorithm 1. The system implements an AI assistant for elderly care, leveraging
conversation capability for user interactions. It also controls a massage subsystem via PWM adjustments
based on foot reflexology principles through a mobile app. Conversation enhances the well-being of elderly
users through personalized care and companionship.

Artificial intelligence driven robotic control system for personalized elderly care … (Shripad Bhatlawande)
42  ISSN: 2089-4864

Algorithm 1. Algorithm for massage control and response generation

Input: Mobile application string(X1) and voice input(X2)
Output: Run foot sole massage(Y1), adjust intensity of calves massage(Y2)
and provide response(Y3)
Initialize:
Device Bluetooth connection
1. Loop:
2. UserInput = X2
3. CharacterInput=X1
4. if X1 is "Running foot sole massage" then
5. Invoke Y1
6. else if CharacterInput is "Running calves massage" then
7. Invoke Y2(5)
8. if IncreaseButtonPressed() then
9. Invoke Y2(5+1)
10. end if
11. else if DecreaseButtonPressed() then
12. Invoke Y2(5-1)
13. end if
14. end if
15. else if X1 is "Stop button pressed" then
16. StopMassage()
17. end if
18. if Prompt(X2) then
19. Emotion= ExtractEmotion(X2)
20. Invoke Y3(Emotion)
21. end if
22. Return

4. RESULTS AND DISCUSSION

The setup under study consists of two separate systems, which are voice assistant and massage
system. The voice assistant takes the input from elderly and processes the input, generating a response based
on the requirements. The foot and calves massage system can also operate through a mobile application,
which enables the user to control the massaging intensity. The whole setup is provided to the users for
testing, some of the responses given by the conversation robot are mentioned below. Eg. user input “Give
medication information related to Azithromycin tablet”, the response to the above input is given as
“Azithromycin is a macrolide antibiotic used to treat bacterial infections. It comes in tablet form, with
common side effects including nausea and diarrhea”. Next input given as “make a reminder at 10 pm today to
take blood pressure medicine” its response “Reminder is set for 10 pm today to take blood pressure
medicine”. Such types of responses are converted to numerical data of how well the AI responded to the
given query, plotted in Figure 4. The standard deviation of the response score is at 1.62 whereas mean of the
feedback data points is 7.70. The emotion classification model was fine-tuned on the foundational BERT-
base-uncase model and trained on the Dair-ai dataset. The validation dataset was used to perform
hyperparameter tuning through various trial and error experiments. After multiple iterations and tests the
below training arguments were found to be the most suitable to achieve high accuracy as mentioned in
Table 3.

Table 3. Hyperparameter values

Hyperparameters Values
Learning rate 2𝑥10−5
Training batch size 64
Evaluation batch size 64
Random seed 42
Optimizer AdamW with betas=(0.9, 0.999) and epsilon=1e-08
LR scheduler type Linear
Number of epochs 5

The model was trained on these optimal hyperparameters and evaluated on the test dataset. The
model after 5 epochs of training achieved an impressive validation accuracy of 93%. The validation loss was
0.1416 at the end of 5 epochs. The model was evaluated on multiple performance metrics such as F1 score,
recall and precision. The effectiveness of the model to correctly predict each of the output classes was
measured using these performance metrics and a classification report was generated as shown in Table 4. The
proposed research article successfully developed an AI driven robotic control system for foot sole and calves’

Int J Reconfigurable & Embedded Syst, Vol. 14, No. 1, March 2025: 35-47
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  43

massage along with conversational bot to provide elderly care and companionship. The prototype was tested
on a sample size of 20 elderly people having ages within the range of 60-85 years and a review of the
massager and the conversational relevance was obtained.

Table 4. Classification performance

Class Precision Recall F1 Score
0 0.96 0.97 0.97
1 0.96 0.94 0.95
2 0.80 0.87 0.83
3 0.93 0.91 0.92
4 0.88 0.92 0.90
5 0.82 0.71 0.76

The reviews of the elderly who used the system, including the foot sole massage and calves’
massage are provided in Figures 4 and 5 respectively. These reviews show the system’s performance in real
life scenarios. The standard deviation in feedback analysis as given in Figure 5 for foot sole massage stands
at 1.98 whereas for conventional foot massage is 2.32. Further the mean feedback rating in foot sole massage
data is 7.10 and in conventional massage data is 5.95 indicating higher rating to the proposed system. The
comparison of the system with the conventional human massage is plotted in Figure 6. The plot portrays that
the presented system has better performance as per the rating given by participants.

Figure 4. Elderly feedback on AI response system (rating out of 10)

Figure 5. Elderly feedback on foot massage system (rating out of 10)

Figure 6. Elderly feedback on calf massage system (rating out of 10)

Artificial intelligence driven robotic control system for personalized elderly care … (Shripad Bhatlawande)
44  ISSN: 2089-4864

The subjective feedback for our system in comparison with conventional leg massage alternatives
available in the market are plotted in Figures 5 and 6 respectively. The standard deviation for calves’
massage and conventional massage are 1.78 and 2.12 respectively. Further, the mean of the data points for
calves’ massage and conventional massage are 7.60 and 5.70 respectively. The trials were conducted on 20
individuals (aged 60 and above). The statistical inference from the plots demonstrates 70% were satisfied
with foot massage and 75% appreciated the calves’ massage over the conventional massage (provided by
humans). The parameters that are taken into consideration for the massage system are: torque applied by
motor on feet, the pressure imparted by cuff on calves and comparison with conventional massage. The
parameter for conversation robot is the relevance of response provided by the system. The evaluation matrix
for massage feedback is represented in Table 5. The matrix depicts the feedback evaluation score provided by
each elderly person for foot sole and calf massage. The matrix also includes a post-modification experience
assessment (PMEA) index, which depicts feedback from the elderly on how the massage system responded to
user input. The index consists of foot sole massage feedback weighted at 60% and calves massage feedback
weighted at 40%. The conversational relevance of the chatbot is evaluated by calculating semantic similarity
between the generated response and a ground truth response. A total of 30 pairs of generated responses and
ground truth responses are collected across all 6 emotion classes. The ground truth responses are manually
curated for each user input. The sentences undergo preprocessing - tokenization, lowercasing, punctuation
removal, and stop words elimination to ensure standardized representations conducive to subsequent analysis.
The pre-processed tokens are transformed into dense vector representations using global vectors (GloVe) for
word representation GloVe word vectors, capturing the semantic nuances inherent in the language. GloVe
word vector representations are renowned for their effectiveness in capturing semantic relationships between
words based on their co-occurrence statistics in large text corpora. The calculated similarity score S is
denoted by (5):

𝑆(𝑖, 𝑗) ∈ [−1,1] (5)

where: 𝑖, 𝑗 denotes the generated response and ground truth response respectively.

Table 5. Massage evaluation matrix

Elderly ID Foot sole evaluation Calves massage evaluation PMEA index
E1 8 8 8
E2 7 8 7.2
E3 5 7 5.8
E4 4 10 6.4
E5 9 8 8.6
E6 4 6 4.8
E7 6 7 6.2
E8 7 10 8.6
E9 8 9 8.6
E10 5 5 5
E11 6 7 6.4
E12 10 10 10
E13 6 4 5.6
E14 10 6 8.4
E15 6 8 6.8
E16 7 5 6.2
E17 9 8.6
E18 8 8 8
E19 7 7 7.2
E20 10 10 10

Values closer to 1 (refer to Table 6) indicate perfect similarity between two vectors implying that
the generated response is ideal and is identical in content and meaning with respect to the ground truth
response. Scores ranging from 0 to -1 indicate complete dissimilarity implying the generated response is
completely out of context or is wrong. The mean similarity score of each sentence pair for each emotion class
is calculated separately as shown in Table 6. A mean similarity score across all emotion classes is calculated
to be 0.79. Results from the experiments reveal several key insights. The foot massage task demonstrated
good performance, with acceptable feedback from the user. Conversation and updates are effective methods
to engage in meaningful conversation with providing relevant information. The similarity scores in Table 6
imply that the conversational bot is highly effective in generating responses and building meaningful
conversations that depict joy, anger or fear while being less relevant where emotions like fear or surprise
come into play.

Int J Reconfigurable & Embedded Syst, Vol. 14, No. 1, March 2025: 35-47
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  45

Table 6. Similarity scores between generated responses and ground truth responses for each emotion class
Emotion Sadness (0) Joy (1) Love (2) Fear (3) Anger (4) Surprise (5)
Semotion 0.84 0.89 0.78 0.73 0.84 0.68

The mobile application’s first page asks for the user to connect with the system using Bluetooth, a
scan button is provided which shows all the visible Bluetooth devices available. When the mobile app is
successfully connected with the system a new window opens which has options to start and stop the foot and
calf massage. The calf massage section contains two more buttons to increase or decrease the intensity of the
calf massage as shown in Figure 7. The system is enabled with a switch for turning the system on/off. It
contains a Rpi mic, which takes the user input and a speaker which delivers the output of the conversation
response to the user is shown in Figure 8. However, it is essential to acknowledge some limitations. The
elderly population's varying responses to the system may affect results in Foot sole and calf massage. The
currently available products in the market for foot and sole massage are significantly expensive, because of
which many people are unable to afford it. Furthermore, the products in the market mainly focus on either
foot massage or calf massage, very few focus on both. The proposed system is cost effective as well as easy
to use and it provides not only companionship but also foot and calf massage to the elderly.

Figure 7. Mobile app GUI

Figure 8. Foot massager system

5. CONCLUSION
The elderly subjects evaluated the system in real and controlled environments. The proposed
research developed an AI Assistant that provides companionship by inducing fruitful conversations,
manipulating massage control systems and also setting precise medication reminders thereby enhancing the
overall user experience and quality of life. The solution stands out in three distinct ways. Firstly, it boasts the
ability to dynamically adapt to individual user preferences, making each massage session personalized to
some extent. Secondly, its real-time responsiveness ensures an uninterrupted experience for the users as
mentioned in the results section. Thirdly, it provides a user-friendly mobile app to control the foot sole
massage and calves massage, with additional features such as controlling the intensity of calves’ massage.
The advantages offered by our system are significant. Users can enjoy a substantial reduction in manual
intervention, leading to a hassle-free experience. Moreover, there's potential for improved therapy outcomes,
making it an invaluable tool for elderly individuals seeking relaxation and wellness.
Artificial intelligence driven robotic control system for personalized elderly care … (Shripad Bhatlawande)
46  ISSN: 2089-4864

It is important to acknowledge a couple of limitations. The effectiveness of our system is contingent

on the quality and diversity of personalized settings for therapeutic goals. In future scope, improvements
could focus on interfacing AI with mobile applications, implementing different modes for massaging and
improving the conversation response. The features to be extended in future work include multiple calf
massage arrangements with an improved pneumatic system. The system currently has limited voice input-
based conversation, to be improved with added functionality of voice assistance.

REFERENCES
[1] A. Pandit, “India aging rapidly, 60+ will make up 15% of population by 2036: UN,” Times of India, Sep. 2023, [Online]
Available: https://timesofindia.indiatimes.com/india/india-aging-rapidly-60-will-make-up-15-of-population-by-2036-
un/articleshow/104001546.cms (accessed Sep. 28, 2023).
[2] Dignity “Addressing loneliness in senior citizens,” Akshardhar, Mar. 2023, [Online] Available:
https://www.akshardhara.org/dignity/addressing-loneliness-in-senior-citizens/ (accessed Mar. 30, 2023).
[3] S. Keshmiri, H. Sumioka, R. Yamazaki, and H. Ishiguro, “Decoding the perceived difficulty of communicated contents by older
people: Toward conversational robot-assistive elderly care,” IEEE Robotics and Automation Letters, vol. 4, no. 4, pp. 3263–3269,
2019, doi: 10.1109/LRA.2019.2925732.
[4] H. Xu, Y. Pan, J. Li, L. Nie, and X. Xu, “Activity recognition method for home-based elderly care service based on random forest
and activity similarity,” IEEE Access, vol. 7, pp. 16217–16225, 2019, doi: 10.1109/ACCESS.2019.2894184.
[5] S. Juraev, A. Ghimire, J. Alikhanov, V. Kakani, and H. Kim, “Exploring human pose estimation and usage of synthetic data for
elderly fall detection in real world surveillance,” IEEE Access, vol. 10, pp. 94249–94261, 2022, doi:
10.1109/ACCESS.2022.3203174.
[6] C. Shang, C.-Y. Chang, G. Chen, S. Zhao, and H. Chen, “BIA: behavior identification algorithm using unsupervised learning
based on sensor data for home elderly,” IEEE Journal of Biomedical and Health Informatics, vol. 24, no. 6, pp. 1589–1600, Jun.
2020, doi: 10.1109/JBHI.2019.2943391.
[7] M. Lee, D. T. Tran, H. Yamazoe, and J. H. Lee, “Care training assistant robot and visual-based feedback for elderly care
education environment,” in 2021 IEEE/SICE International Symposium on System Integration, SII 2021, 2021, pp. 572–577, doi:
10.1109/IEEECONF49454.2021.9382630.
[8] A. Y. Alaoui, S. El Fkihi, and R. O. H. Thami, “Fall detection for elderly people using the variation of key points of human
skeleton,” IEEE Access, vol. 7, pp. 154786–154795, 2019, doi: 10.1109/ACCESS.2019.2946522.
[9] W. Si, G. Srivastava, Y. Zhang, and L. Jiang, “Green internet of things application of a medical massage robot with system
interruption,” IEEE Access, vol. 7, pp. 127066–127077, 2019, doi: 10.1109/ACCESS.2019.2939502.
[10] M. Nasr, M. M. Islam, S. Shehata, F. Karray, and Y. Quintana, “Smart healthcare in the age of AI: recent advances, challenges,
and future prospects,” IEEE Access, vol. 9, pp. 145248–145270, 2021, doi: 10.1109/ACCESS.2021.3118960.
[11] E. Martinez-Martin and A. Costa, “Assistive technology for elderly care: an overview,” IEEE Access, vol. 9, pp. 92420–92430,
2021, doi: 10.1109/ACCESS.2021.3092407.
[12] Z. Zhu et al., “CBASH: a CareBot-assisted smart home system architecture to support aging-in-place,” IEEE Access, vol. 11, pp.
33542–33553, 2023, doi: 10.1109/ACCESS.2023.3264272.
[13] Y. Wei, K. Gu, X. Cui, C. Hao, H. Wang, and Y. Chang, “Strategies for feet massage robot to position the pelma acupoints with
model predictive and real-time optimization,” International Journal of Control, Automation and Systems, vol. 14, no. 2, pp. 628–
636, 2016, doi: 10.1007/s12555-014-0273-3.
[14] A. Paul et al., “Health efficacy of electrically operated automated massage on muscle properties, peripheral circulation, and
physio-psychological variables: a narrative review,” Eurasip Journal on Advances in Signal Processing, vol. 2021, no. 1, 2021,
doi: 10.1186/s13634-021-00788-6.
[15] E. G. Christoforou, A. S. Panayides, S. Avgousti, P. Masouras, and C. S. Pattichis, “An overview of assistive robotics and
technologies for elderly care,” in IFMBE Proceedings, 2020, vol. 76, pp. 971–976, doi: 10.1007/978-3-030-31635-8_118.
[16] G. Bardaro, A. Antonini, and E. Motta, “robots for elderly care in the home: a landscape analysis and co-design toolkit,”
International Journal of Social Robotics, vol. 14, no. 3, pp. 657–681, 2022, doi: 10.1007/s12369-021-00816-3.
[17] M. Niemelä and H. Melkas, “Robots as social and physical assistants in elderly care,” in Human-Centered Digitalization and
Services, 2019, pp. 177–197, doi: 10.1007/978-981-13-7725-9_10.
[18] A. A. Khan, C. Wechtaisong, F. A. Khan, and N. Ahmad, “A cost-efficient environment monitoring robotic vehicle for smart
industries,” Computers, Materials and Continua, vol. 71, no. 1, pp. 473–487, 2022, doi: 10.32604/cmc.2022.020903.
[19] A. A. Khan and F. A. Khan, “A cost-efficient radiation monitoring system for nuclear sites: designing and implementation,”
Intelligent Automation and Soft Computing, vol. 32, no. 3, pp. 1357–1367, 2022, doi: 10.32604/IASC.2022.022958.
[20] D. Zhao et al., “Research status of elderly-care robots and safe human-robot interaction methods,” Frontiers in Neuroscience, vol.
17, 2023, doi: 10.3389/fnins.2023.1291682.
[21] R. Y. S. Kopp et al., “Conversational assistants for elderly users–the importance of socially cooperative dialogue,” in Proceedings
of the AAMAS Workshop on Intelligent Conversation Agents in Home and Geriatric Care Applications co-located with the
Federated AI Meeting, 2018, pp. 10–17.
[22] M. Zhu et al., “A peristaltic soft, wearable robot for compression therapy and massage,” IEEE Robotics and Automation Letters,
vol. 8, no. 8, pp. 4665–4672, 2023, doi: 10.1109/LRA.2023.3287773.
[23] W. Hu, Q. Sheng, and X. Sheng, “A novel realtime vision-based acupoint estimation for TCM massage robot,” in 2021 27th
International Conference on Mechatronics and Machine Vision in Practice, M2VIP 2021, 2021, pp. 771–776, doi:
10.1109/M2VIP49856.2021.9665080.
[24] S. N. Sayapin, “Decision-making by the autonomous symbiotic self-relocating massage robot ‘octahedral dodekapod’ based on
SEMS during the upper or lower limb massage,” in Studies in Systems, Decision and Control, vol. 352, 2021, pp. 53–70, doi:
10.1007/978-3-030-68172-2_5.
[25] Y. Kerautret, F. Di Rienzo, C. Eyssautier, and A. Guillot, “Comparative efficacy of robotic and manual massage interventions on
performance and well-being: a randomized crossover trial,” Sports Health, vol. 16, no. 4, pp. 650–660, 2024, doi:
10.1177/19417381231190869.

Int J Reconfigurable & Embedded Syst, Vol. 14, No. 1, March 2025: 35-47
Int J Reconfigurable & Embedded Syst ISSN: 2089-4864  47

BIOGRAPHIES OF AUTHORS

Shripad Bhatlawande received Ph.D. degree from the Indian Institute of

Technology Kharagpur, India, in 2015. He received bachelor’s degree in electronics
engineering from the SGGS COE, Nanded, India, in 2000 and master’s degree in electronics
engineering from the Government College of Engineering, Pune, India, in 2008. His research
interests include embedded systems, machine intelligence, and robotics. He can be contacted at
email: [email protected].

Swati Shilaskar pursued Ph.D. From Government College of Engineering

Amravati. She received the B.E. degree in electronics engineering, the M.E. degree in digital
electronics from Sant Gadge Baba Amravati University, India. Her research interests include
VLSI design, BCI, medical diagnostic support systems, and automation. She can be contacted
at email: [email protected].

Soham Akotkar is a dedicated student pursuing a B.Tech. in electronics and

telecommunication at the esteemed Vishwakarma Institute of Technology, an autonomous
institute affiliated with Savitribai Phule Pune University. His academic interests center around
the cutting-edge fields of embedded systems and robotics. His commitment to innovation is
evident in his current project, "AI assisted robotic system". His previous projects include
Railway track health monitoring robot and hardware trojan detection in power devices and
embedded systems. He has submitted a research paper on railway track health monitoring
project to CISCON (Control Instrumentation Systems conference) for consideration. He has
established himself as a runner up in the smart India hackathon 2023 (hardware edition). He
can be contacted at email: [email protected].

Anish Joshi is an AI enthusiast pursuing B.Tech. in electronics and

telecommunication engineering in Vishwakarma Institute of Technology, an autonomous
institute affiliated to Savitribai Phule Pune University. His research interests mainly focus on
deep learning in computer vision and natural language processing tasks. He currently works as
deep learning intern and contributes to a team of machine learning engineers that develop,
optimize and deploy deep learning models in a remote sensing environment. His previous
research work “Structural health monitoring of railway track: RailBot” was accepted in the
Control Instrumental Systems Conference (CISCON 2023). He was instrumental in developing
a novel architecture for crack detection using CNN. He can be contacted at email:
[email protected].

Zayd Ansari is an enthusiastic student pursuing a B.Tech. in electronics and

telecommunication at Vishwakarma Institute of Technology, an autonomous institute affiliated
with Savitribai Phule Pune University. His research focuses on the elderly care domain for
physical aid, with his previous project, "Renewable energy trading platform using blockchain''
demonstrating his innovative approach to real-world problems. He has submitted a research
paper on this project to International Conference on Innovative Data Communication
Technologies and Application (ICIDCA) for consideration. Beyond his academic pursuits,
Zayd has established himself as a runner up in the smart India hackathon 2023 hardware
edition. He can be contacted at email: [email protected].

Artificial intelligence driven robotic control system for personalized elderly care … (Shripad Bhatlawande)