Low-Power Smart Home Automation via

On-Device Gesture Recognition Using Nano BLE

Mr. Yash Tagunde
MIT School of Computing,
MIT ADT University
Pune, India
[email protected]
Mr. Shrishant Madke
MIT School of Computing,
MIT ADT University
Pune, India
[email protected]

Prof. Harshad Lokhande

MIT School of Computing,
MIT ADT University
Pune, India

Mr. Vedant Patil

MIT School of Computing,
MIT ADT University
Pune, India
[email protected]
Mr. Prem Urkude
MIT School of Computing,
MIT ADT University
Pune, India
[email protected]

1
 Abstract Impulse. The other hardware components used are jumper cables,
LED bulbs, DC motor fan.
This paper presents a smart home automation system that
utilizes gesture recognition powered by edge machine learning B. Data Collection and Labeling
on the Arduino Nano BLE Sense microcontroller. The system is Motion data was collected for four gesture categories:
designed to detect hand gestures such as up, down, left, right,  Idle (25 samples)
idle and more which are used to control home appliances like
lights and fans. The model is trained using Edge Impulse to  Circle (25 samples)
recognize gesture patterns based on accelerometer and
 Right-Left (25 samples)
gyroscope data from the Nano BLE Sense. Unlike traditional
systems that rely on voice commands or smartphone  Up-Down (25 samples)
applications, the proposed approach provides a contactless,
Each sample recorded 4000 ms of IMU data at a frequency of
efficient, and privacy-friendly solution for home automation,
100 Hz, resulting in a total of 94 training samples and 27 testing
especially beneficial for elderly or differently-abled individuals. samples (collected similarly). The total duration of all captured data
Experimental results demonstrate high gesture recognition amounts to 8 minutes and 10 seconds. The dataset was split into
100% accuracy and real-time control responsiveness, 77% for training and 23% for testing.
validating the feasibility of deploying ML models on low-power
embedded devices for smart living applications. C. Impulse Design and Model Training
Keywords: Smart Home Automation, Gesture Detection, Edge Using Edge Impulse, a time series classification model was
Machine Learning, Arduino Nano BLE Sense, Edge Impulse, created. The impulse design involved:
Contactless Control  Window Size: 2000 ms
 Frequency: 100 Hz
I. INTRODUCTION  Features Extracted: Spectral and time-domain features
The evolution of smart homes has been significantly driven by the from accelerometer and gyroscope axes.
integration of automation and artificial intelligence, offering On-device performance:
enhanced comfort, security, and energy efficiency. Conventional  Peak RAM usage: 3 KB
smart home automation systems often rely on voice commands or  Inference time: 34 ms
mobile applications, which may not always be accessible or D. Deployment and Real-World Testing
intuitive for all users, particularly for the elderly or differently-abled
individuals. In light of these challenges, gesture-based control After training, the model was deployed directly onto the Arduino
systems have emerged as a compelling alternative, offering a hands- Nano BLE Sense. A test environment mimicking a smart home was
free and natural method of interaction. created using a breadboard circuit where:
Recent advancements in edge computing and embedded machine
learning have made it possible to deploy intelligent systems on low-
 A LED bulb represented a room light.
power microcontrollers. This paper proposes a low power smart  A DC motor simulated a fan.
home automation solution that utilizes gesture detection through
machine learning deployed on the The control logic:
Arduino Nano BLE Sense microcontroller. By training a model to
classify specific hand gestures, such as up-down,  Right-Left gesture → Turn ON the light.
 Up-Down gesture → Turn OFF the light.
right-left, circular, and idle, the system can control household
appliances like lights and fans without any physical interaction.
 Circle gesture → Toggle the fan ON/OFF.
 Idle gesture → No change in appliance state.
Edge Impulse, a leading platform for embedded ML development, is
used to collect and label motion data, train a time series The gesture inference and appliance control program was
classification model, and deploy the model directly onto the device.
The proposed solution aims to deliver a real- written in Arduino IDE using the generated Edge Impulse
inference library. The system responded to each gesture in
real time with high reliability, validating its effectiveness in
a real-world context.

time, cost-effective, and privacy-friendly approach to smart home IV. RESULTS AND DISCUSSION
automation, reducing dependence on cloud services and external To thoroughly assess the viability of the proposed smart
devices. The remainder of this paper details the data collection,
model training, deployment strategy, real-world implementation,
home automation system using gesture detection, a series of
and evaluation metrics of the system. structured experiments were conducted. The system’s
performance was evaluated across multiple dimensions,
II. LITERATURE SURVEY including model accuracy, memory efficiency, processing
time, and real-time responsiveness when deployed on the
Arduino Nano BLE Sense. Both the training and testing
phases involved gesture data collection over a defined
III. METHODOLOGY duration, with precise labeling and segmentation of gesture
classes. Furthermore, the implementation was tested in a
A. Hardware Components real-world setup where gestures triggered physical
The primary hardware used is the Arduino Nano 33 BLE Sense components like a fan and an LED bulb. The results of these
microcontroller, equipped with a 9-axis Inertial Measurement Unit evaluations, as well as insights drawn from performance
(IMU), which includes an accelerometer and gyroscope. It is chosen metrics and live tests, are discussed in the following section.
for its compact size, BLE capabilities, and compatibility with Edge
A. Circuit Diagram :- V. CONCLUSION
This research presents a low-cost, real-time, and intuitive
gesture-based smart home automation system using the
Arduino Nano BLE Sense and Edge Impulse platform. By
accurately recognizing hand gestures, the system effectively
controls appliances like lights and fans without the need for
physical switches or voice commands.
With excellent model performance and smooth on-device
operation, this project showcases the potential of embedded
machine learning for practical home automation solutions.
Future advancements could focus on user adaptability,
gesture diversity, and expanding functionality to
accommodate broader smart home ecosystems.

Figure 1
A. Figures and Tables
B. Model Performance
The trained model demonstrated excellent performance
metrics during testing and validation. With an accuracy of
100%, and evaluation metrics such as precision, recall, F1-
score, and area under the ROC curve all equal to 1.00, the
model is highly robust and reliable. This indicates that the
gesture classes are well-separated and the classifier is
consistent across different gestures.
C. On-Device Testing
The on-device inference performance was impressive, with a Figure 2
minimal RAM usage of 3 KB and an average processing
time of 34 ms per inference. This enables near real-time
response from the system. The microcontroller was able to ACKNOWLEDGMENT
handle gesture classification and appliance control
simultaneously without any significant delay or error. I would like to extend my heartfelt gratitude to Prof.
D. Real-World Usability Harshad Lokhande for his invaluable guidance, mentorship,
In the simulated smart home environment, the system and constant encouragement throughout the course of this
responded accurately to gesture inputs. For instance: project. His insights, technical expertise, and constructive
feedback were instrumental in shaping the direction and
 Performing a right-left gesture consistently turned outcome of this research. I am also thankful to the faculty
the LED light on. and staff of the Department of Computer Engineering at
MIT ADT University for providing a supportive academic
 An up-down gesture reliably turned the light off. environment and access to necessary resources.

 Circular gestures toggled the fan state between ON Additionally, I would like to acknowledge the creators and
and OFF effectively. contributors of the Edge Impulse platform and the Arduino
Nano BLE Sense community, whose documentation and
 The system maintained the last state when in idle, tools significantly facilitated the development and
avoiding unnecessary switching. deployment of the machine learning model. Last but not
least, I express my gratitude to my peers, friends, and family
This performance shows the feasibility of using gesture- for their unwavering support and motivation throughout this
based systems in real-world smart home scenarios, journey.
especially where traditional controls are impractical.
REFERENCES
E. The model achieved the following training performance:
[1] Edge Impulse. (n.d.). The leading development platform for
embedded machine learning. Retrieved from
 Accuracy: 100% https://www.edgeimpulse.com
 Loss: 0.00
[2] Arduino. (n.d.). Arduino Nano 33 BLE Sense. Retrieved from
 Precision: 1.00 https://docs.arduino.cc/hardware/nano-33-ble-sense
 Recall: 1.00
 F1 Score: 1.00 [3] Arduino. (n.d.). Arduino IDE. Retrieved from
 Area under ROC Curve: 1.00 https://www.arduino.cc/en/software

[4] TensorFlow Lite for Microcontrollers. (n.d.). Deploy machine

learning models on microcontrollers and edge devices. Retrieved
from https://www.tensorflow.org/lite/microcontrollers
 [5] K. Simonyan & A. Zisserman. (2014). Two-Stream focus on increasing the dataset size and diversity to improve
Convolutional Networks for Action Recognition in Videos. In
the robustness and generalization capability of the gesture
Advances in Neural Information Processing Systems (NeurIPS).
recognition model, particularly under varying lighting
[6] M. R. Endsley. (1995). Toward a Theory of Situation Awareness conditions, hand orientations, and user hand sizes.
in Dynamic Systems. Human Factors, 37(1), 32–64. Additionally, expanding the range of recognized gestures
[7] P. Corke. (2017). Robotics, Vision and Control: Fundamental
could enable more granular control of home appliances and
Algorithms In MATLAB. Springer. introduce advanced features like dimming lights, adjusting
fan speeds, or triggering emergency responses. Integration
[8] Scikit-learn Developers. (2011). Scikit-learn: Machine Learning with other IoT protocols such as MQTT and platforms like
in Python. Journal of Machine Learning Research, 12, 2825–
2830. Retrieved from https://scikit-learn.org
Blynk or Home Assistant could enhance remote monitoring
and multi-device coordination capabilities.
[9] D. Zhang, Y. Ma, & Y. Lu. (2017). Smart Home: Architecture, Another potential area of development is the incorporation
Technologies, and Systems. Procedia Computer Science, 131, of low-power communication modules, such as Bluetooth
250–257.
Low Energy (BLE) or Wi-Fi, to enable seamless interaction
[10] M. Hassan, B. Song, & M. Hossain. (2015). A framework of with other smart devices. Real-time performance
smart home system based on machine learning algorithms. benchmarking under different deployment scenarios, such as
Computers & Electrical Engineering, 52, 276–288. battery-powered operation and multiple room settings, will
be investigated.
Future Work
Lastly, user-centric enhancements such as adaptive gesture
While the proposed gesture-based smart home automation learning based on individual usage patterns and
system has demonstrated promising results in controlled personalization of controls could make the system more
conditions, several opportunities exist for further intuitive and inclusive, especially for the elderly and
enhancement and broader applicability. Future work will differently-abled users.