2023 International Conference on Applied Intelligence and Sustainable Computing (ICAISC)

AI Based Robot for Beach Cleaning

1st Ashitha V Naik 2nd E Vivin Raj 3rd Chaitra C T
Dept. of Electronics and Dept. of Electronics and Dept. of Electronics and
Communication Engineering Communication Engineering Communication Engineering
Nitte Meenakshi Institute of Nitte Meenakshi Institute of Nitte Meenakshi Institute of
2023 International Conference on Applied Intelligence and Sustainable Computing (ICAISC) | 979-8-3503-2379-5/23/$31.00 ©2023 IEEE | DOI: 10.1109/ICAISC58445.2023.10200181

Technology Technology Technology

Bangalore, India Bangalore, India Bangalore, India
[email protected] [email protected] [email protected]

4th Harshitha K S 5th Komal

Dept. of Electronics and Dept. of Electronics and
Communication Engineering Communication Engineering
Nitte Meenakshi Institute of Nitte Meenakshi Institute of
Technology Technology
Bangalore, India Bangalore, India
[email protected] [email protected]

Abstract—“Beach cleaning robots” are used to pick up trash Due to the increasing sewage issues from nearby industry,
and debris from beaches. Beach cleaning robots are becoming it is urgent to address the sewage water issue as businesses
more and more popular since they can work continuously and continue to expand. Industrial pollutants are extremely
effectively without human labor, keeping beaches clean and safe hazardous to both humans and the environment. The second
for tourists. These robots can assist in reducing the quantity of most essential element is a waste management system, which
plastic and other rubbish that ends up in the ocean and can also allows workers to preserve their health and productivity by
assist in protecting wildlife from harm caused by beach litter. ensuring that reporting workers do not have to wait and enter
The Raspberry Pi is a crucial component in our work to drainage. Fig 1 shows the beach is littered with garbage.
construct the AI-based robot. The Raspberry Pi Pico is used to
Beach cleaning is a significant issue these days. NGOs and
control the movement of the object in the photo that is taken
with the Pi- Camera. The waste is identified, using motors and
other non-profit organizations are working tirelessly to keep
a conveyor belt; it is collected and dumped separately to beaches clean. However, cleaning beaches by hand requires a
maintain cleanliness. significant amount of workforce, volunteers, and enthusiasm.
This causes several environmental issues. Prohibited beach
Keywords—Raspberry Pi, Raspberry Pi Pico, Tensor Flow, use and reduced circulation in public spaces produced a
OpenCV, Pi Camera Interfacing, Image Processing. decrease of up to 83% in the total amount of litter on the
beaches studied [1].
I. INTRODUCTION
To prevent this and conserve our time control of waste, we
The most serious problem in countries such as India, that are developing the "Smart cleaning system application”. As
are renowned for their beaches, occurs when people the world becomes more digitalized and automated, today's
overcrowd and begin trashing previously cleaned beach, kids want everything to be simple and smart. People of all ages
compromising the beauty of those tourist destinations. are finding it extremely simple to be smart effort and to be
Beaches are a major attraction for tourists. Our capacity to use healthier more frequently.
beaches for commercial, recreational, and aesthetic reasons is
constrained by the contamination of the coastal environment. II. RELATED WORKS
It further deteriorates and obliterates the special beach The RF machine learning classifier was able to operate on
ecosystem that is necessary for living creatures and plants. We a cheap Raspberry Pi 4 with the UASs under test, providing
have developed cleaning equipment that enables us to clean an overall accuracy for a two-class recognition system of
beaches in a shorter amount of time since cleaning machines 100% and UAS type classification accuracy of 90.9%. The
must be utilized to clean the beach. design incorporates elements from various initiatives all over
the globe, such as a chain sprocket setup from one project and
the structure of another, a trash picking technique from one
project and a trash deposition method from another, resulting
in an innovative growth combo of a number of beach cleaner
devices which have been invented and crafted in the past [2].
The primary goal of this initiative is to help with pollution
issues causing harm the environment. The HS-Green Fist
robot's functionality is to be able to move and gather bottles in
a certain region before transporting the gathered cans to a
deposit.The ESP32 CAM board was used to power the BCR,
which also includes sensors for obstacle detection and
automatic garbage collection. The ESP32 serves as an
interface between the BCR, and the web server [3].
To achieve accurate and real-time trash recognition, the
Fig. 1. Beach littered with garbage. YOLOv3 network, which is extensively used in the field of

979-8-3503-2379-5/23/$31.00 ©2023 IEEE

Authorized licensed use limited to: College of Engineering - THIRUVANANTHAPURAM. Downloaded on August 17,2023 at 08:53:18 UTC from IEEE Xplore. Restrictions apply.
 high speed and accuracy object detection, is trained on the three components of the mapping robot that are discussed in
suggested floating rubbish dataset. To discriminate rubbish this article. Each chassis motor (up to six in total) is coupled
while the cleaning robot is moving, the object identification to a platform that offers more movement angles, enabling the
algorithm should be precise and real-time [4]. vehicle to traverse uneven terrain. In this project, the chassis
is linked to four motors, creating four wheels. Each motor is
The suggested Beach Litter Assessment Technique controlled by two bits, which determine whether it spins
Qualitative and Quantitative method's subjective component clockwise or anticlockwise. [11].
recognized the primary categories of beach litter, while the
numeric segment provided a summary of the level of beach III. METHODOLOGY
hygiene. Machine learning models extract insight from data
by utilizing algorithms based on frameworks such as Tensor The main system of the AI Based Robot for Beach Cleaning
Flow. During the training cycle, models learn. Python is the isdivided into two subsystems as shown in the Fig 2. Each
recommended language of programming for Tensor Flow subsystem includes stages of both hardware construction and
model development [5]. software programming. For the system architecture of this
project,the suggested technique is to have a core system that
The paper presents a methodology which identifies the will help with the cleaning process. The machine will include
issues in maintaining repositories for TensorFlow machine four wheels, a container that will hold most of the machine's
learning projects. In this paper it suggests that ML specific parts, and a conveyor belt for rubbish collecting with hooks
platforms like MLflow, Polyaxon and Comet.ml are better for attached. The container-like design of the robot's construction
maintaining repositories of ML projects and to control shields it from sand and other items. The device contains a
repeating filtration of the model building process. Repository Raspberry Pi, the Microprocessor that controls the robot. In
doesn’t allow the user selection of context information. These addition, the Raspberry Pi will control a camera and transmit
technologies are unable to be used for projects in which ML data to a remote server for evaluation.The robot will detect
features are added gradually to conventional software since garbage by using its in-built camera module.
they do not permit retaining repositories of non-ML projects.
[6]. The block diagram of the AI Based for Beach Cleaning is
as shown in fig 3. The main part of this project is the
To maneuverer the hTetro robot from qs to qs+1, the best- Microprocessor the Raspberry pi. The Lithium-Ion battery is
performing motion command sequence must be found using connected to the robot to give power supply. All the Bo
the GA (Genetic Algorithm) operation in each setup loop. Motors and Servo Motor are connected to the H Bridge Motor
The MOOP uses the following criteria to assess the Driver for the movement of the robot. The robot will detect
effectiveness of the robot motion command: (1) goal garbage by using its in-built camera module. The Ultrasonic
attainability, (2) time consumption, (3) path smoothness, and sensor is used for measuring the distance and for detection
(4) path safety [7]. purposes while the robot is in action and detecting litter on the
beach. The robotic vision system’s image quality can be
In order to improve the neural network and create the
impacted by a range of environmental factors, including but
most suitable framework network pattern to handle issues in
not limited to lighting conditions, shadows, reflections, and
the best way, this study emphasizes the use of optimization
weather conditions like rain and fog [12].
algorithms to modify the tuning parameters or training
variables of the neural network. Using PSO, GA, ABC, and
BSA optimization approaches, some results for enhancing
ANN performance are included in this study [8].
For this objective, a prototype of an RC beach cleaning bot
was created. It has a driving mechanism and a filtering device.
The filtering mechanism consists of a device that uses
vibration to separate small items of trash (such as plastic,
glass, cans, and cigarette butts) from sand. While the driving
mechanism operates the robot's mobility remotely using a
Bluetooth module to transmit and receive signals using a
transceiver. [9].
The ability to identify rubbish is essential for garbage
collection. The robot won't function correctly if its recognition
module misidentifies the garbage or mistakenly interprets a
non-garbage item as rubbish. So, the recognition precision
was checked out first. The dataset has six classes—five trash
classes and one non- garbage class—with a total of 40 000
training photos and 7 000 testing images. Given the absence
of visual characteristics, the mistake in waste paper is
undoubtedly larger than the error in other categories. Similar Fig. 2. System Architecture
to how transparent items have fewer characteristics and are
harder for deep neural networks to recognize, some thin
plastic bags cause the mistake of the bag's contents to be rather
high. [10].
The chassis, on which the tires and their motors are
mounted, an Arduino board, and the power supply are the

Authorized licensed use limited to: College of Engineering - THIRUVANANTHAPURAM. Downloaded on August 17,2023 at 08:53:18 UTC from IEEE Xplore. Restrictions apply.
 analysis between the embedded user and product data, a
multilayer perceptron, and an output layer [14].
1. Convolutional layer: - Convolution occurs when a
kernel or filter within this layer moves across the
image's receptive fields.
2. Pooling layer: - It reduces the dimension of the image
pixels, sacrificing information loss. On the plus side,
this layer decreases complexity while increasing
efficiency.
3. Fully connected layer: - The FC layer in the CNN is
whereimage classification occurs based on the features
derived in the previous layer.
Fig. 3. Block diagram of Proposed System

A. Pi Camera interfacing
Plug the Camera Module ribbon cable, assuring the
connectors at the bottom face the contacts in the port. Return
the plastic clip to its original position. After attaching the PI
Camera module, The Raspberry PI software can be installed
from its official website.
Download the Raspberry PI software and then the VNC
Viewer from the Raspberry PI website. After getting the IP
address of the Raspberry Pi, VNC viewer is used connect the
laptop with the Raspberry Pi.
Fig. 5. Input image and its channel representation.
Then starting up the Raspberry Pi and opening the
Raspberry Configuration tool for configuring the Pi. We As shown in the fig 5, each picture on an electronic gadget
create a file for the python program and enter the python is stored as a channel, which is an arrangement of pixel values.
program for controlling the camera in Thonny IDE. After this Each pixel's value is expressed in eight bits. For each channel
running, the program will ensure the activation of the Pi (RGB), the pixel valueswould then range from 0 to 255. Here
Camera. A video recording will be saved, and it can be viewed we make use of kernels (afeature detector). The input picture
on a screen. This is called PiCamera interfacing. is divided in half, the dot result of the two matrixes is
computed, and the outcomes are saved to a new matrix.
B. Image Processing Pooling layers are used to down sample feature maps,
Artificial Intelligence (AI) has significantly impacted preserving the most relevant sections, and rejecting the rest,
technological advancements in various fields, such as visual reducing the image's spatial size. The most frequent sort of
surveillance, predictive maintenance, object detection, and pooling is Max pooling to get the most important feature of
image segmentation. Specifically, due to computing the image. This section of CNN is utilized for feature
advancements and a growing focus on efficient machine extraction to detect high- level characteristics with as little
learning methods, the scientific community has made great spatial resolution as feasible. The next part is called a
strides in the 3D deep learning domain [13]. Convolutional classifier. This part consists of fully connected layers, like a
Neural Networks (CNN) are deep learning algorithms that are feed forward network. With a sufficient number of neurons
used for image recognition and processing. It is made up of being trained with raw images, we can use this layer for image
convolutional layers, pooling layers, and fully connected classification.
layers. Fig 4. Shows the basic convolutional neural network
OpenCV (Open-Source Computer Vision Library) is a
architecture.
softwarelibrary for computer vision and machine learning that
is used to inputimages. And execute a few transformations on
the raw image to an RGB tensor model in the model format.
Tensor Flow features an image recognition tool, and these
photos are saved in a dedicated folder. When developing AI
models, it is important to consider metrics that focus on the
performance and efficiency of the model. Efficiency is highly
dependent on high detection rates and low false positive or
false negative rates [15].
As shown in the fig 6. training the data using tensorflow
Fig. 4. Convolution neural architecture and OpenCV is needed. Then testing the model by using the
PI camera is necessary to detectthe image and feed an input
Easier patterns like curved lines are identified initially by image of the trash. Once the image is detected, the features are
thelayers, then more complex patterns like faces and objects. extracted out from the detected object. Thenthe features of the
The proposed model comprises an input layer for inputting object are compared with the features of the pre- trained
and embedding the product and user data, a feature extraction model. Once comparison is completed, it shows the output as
layer for extracting the features through data correlation

Authorized licensed use limited to: College of Engineering - THIRUVANANTHAPURAM. Downloaded on August 17,2023 at 08:53:18 UTC from IEEE Xplore. Restrictions apply.
 shown in the fig 7,whether the object can be considered a trash
or not depending on thetrained samples.

Fig. 8. Beach cleaning model

Fig. 6. Training of the model
The robot moves and when it approaches an object, the pi
camera captures it and extracts the features of the object.
These features are then compared with the features present
inside the model. Once the object gets detected and considered
as Trash, as shown in below figure with the percentage rate
which tells us about the accuracy of identifying the object. The
robot moves towards it closer, and the conveyor belt rotates.
As the conveyor belt rotates, the trash gets collected using the
L-shaped hook's peasant on the conveyor belt and throws the
trash in the bin present behind the robot

Fig. 7. Bottle has been recognised.

IV. RESULTS
MDF (Medium density fibreboard) is the material used for
the model as shown in the fig 8. MDF is a preferred material
over wood as it is easier to paint, lighter in weight and can
withstand the weights of the components is being used. The
Servo motor controls the rotation of the Ultrasonic sensors
well as the angle and rotation speed. Using an ultrasonic
sensor, the model (a robot) was able to detect the existence of
an obstruction in its path. The user must verify if the system
is on or off. Supplying the power to the system in the form of Fig. 9. Performance Graph
lithium-ion battery will ensure the systemmoves forward. The
ultrasonic sensor transmits the echo waves and receives the The graph in the fig 9 shows how well the model can make
bounced back signal, if there is an object present in the region decisions by comparing the on pre-trained values added to it
of the transmitted wave. If the distance is more than 20 cm, it with the values acquired from the test picture. It takes the input
continues to move forward. Else if the distance is less than values (x,y,z) according to the coordinate system and
20cm, it stops and moves a few cm backwards. Then the servo compares the same coordinates with the graph above and
motor controls the rotation and the angle of the ultrasonic makes the decision. Table1 and graph in the fig 10 gives the
sensor and turns it towardsright within the 45° angle. It reads information about comparison of image processing techniques
the distance in the right-side direction. Then the sensor turns based on accuracy and sensitivity from various existing
left and measures the distance and sends the details to the methods.
Raspberry Pi. The Raspberry Pi decides whether the right
distance is greater than the left distance. The model's TABLE I. COMPARISON WITH EXISTING METHODS
movement is controlled by the signal given by the H- Bridge Method Accuracy Sensitivity Fi-score
Module from the Raspberry Pi Pico Proposed model 89.4% 87.93% 0.96
Shihan Kong, et al. [3] 87.9% 86.55% -----
PS Janardhanan [5] 87.45% 89.33% -----

Authorized licensed use limited to: College of Engineering - THIRUVANANTHAPURAM. Downloaded on August 17,2023 at 08:53:18 UTC from IEEE Xplore. Restrictions apply.
 [9] Nasreen Bano,Amina Amin, "Radio Controlled Beach Cleaning Bot
",2019 6th IEEE International Conference on Engineering
Technologies and Applied Sciences (ICETAS).
[10] JinqiangBai, ShiguoLian, Zhaoxiang Liu, Kai Wang , and Dijun Liu
"Deep Learning Based Robot for Automatically Picking Up Garbage
on the Grass" IEEE Transactions On Consumer Electronics, Vol. 64,
No. 3, August 2020.
[11] Tunai P. Marques, Fumio Hamano "Autonomous robot for mapping
using ultrasonic sensors" 2019 IEEE Green Energy and Smart Systems
Conference (IGESSC).
[12] Renu Popli, Isha Kansal, Jyoti Verma, Vikas Khullar , Rajeev Kumar
and Ashutosh Sharma “ROAD: Robotics-Assisted Onsite Data
Collection and Deep Learning Enabled Robotic Vision System for
Identification of Cracks on Diverse Surfaces”, Sustainability 2023, 15,
9314, June 9, 2023, MDPI.
[13] Jie Wang, Richard Chang, Ziyuan Zhao and Ramanpreet Singh Pahwa
“Robust Detection, Segmentation, and Metrology of High Bandwidth
Memory 3D Scans Using an Improved Semi-Supervised Deep
Learning Approach”, Sensors 2023, 23, 5470, June 2023, MDPI
[14] Ji-Yoon Kim and Chae-Kwan Lim, “Feature Extracted Deep Neural
Collaborative Filtering for E-Book Service Recommendations”, Appl.
Fig. 10. Graph for Comparison Table Sci. 2023, 13, 6833, June 2023, MDPI
[15] Martin Chovanec, Martin Hasin, Martin Havrilla and Eva Chovancová
V. CONCLUSION AND FUTURE SCOPE “Detection of HTTP DDoS Attacks Using NFStream and
TensorFlow”, Appl. Sci. 2023, 13, 6671, May 2023, MDPI.
This robot is anticipated to solve the rubbish problem,
particularlyon the beach for real time applications. Obstacle
detection using ultrasonic sensors and movement of the robot
as well as object detection using image processing is done.
Accuracyand overall efficiency have been improved by using
raspberry pi microcontroller. The model is easy to handle and
autonomous movement of robot, which makes it an efficient
robot among others. Further in upcoming days after enhancing
each feature and learning about more features, it can be
implemented on largescale. The model is easy to handle and
less costly that makes it an efficient robot among others.
Further in the upcoming days after enhancing each feature, it
can be implemented on a large scale. It can be made
waterproof so that they can avoid the water damaging the
components of the robot.
REFERENCES
[1] José R. Souza Filho, Antônio A. S. Chagas, Iracema R. Silva, Junia K.
Guimarães, Tania E. Sakanaka and Gerson Fernandino “Litter
Reduction during Beach Closure in the Context of the COVID-19
Pandemic: Quantifying the Impact of Users on Beach Litter
Generation”, Pandemic: Quantifying the Impact of Users on Beach
Litter Generation. Sustainability, Januray 2023, MDPI.
[2] Carolyn J. Swinney, John C. Woods, “Low-Cost Raspberry- Pi-Based
UAS Detection and Classification System Using Machine Learning”,
MDPI Aerospace journal ,November 2022
[3] Hadiya Ebrahim, Wahaj Sheikh, Atif Saeed, “Design And Analysis Of
Sustainable Beach Cleaner” MDPI Aerospace journal, November
2022.
[4] Shihan Kong, Manjun Tian, Changlin Qiu, Zhengxing Wu, Junzhi Yu
"IWSCR: An Intelligent Water Surface Cleaner Robot for Collecting
Floating Garbage" IEEE Transactions onsystems, man and cybernetics:
systems, vol. 51, NO. 10,October 2021.
[5] Camilo M. Botero, Diana Tamayo, Seweryn Zielinski and Giorgio
Anfuso, “Qualitative and Quantitative Beach Cleanliness Assessment
to Support Marine LitterManagement in Tropical Destinations” MDPI
Aerospace journal, December 2021.
[6] PS Janardhanan, “Project repositories for machine learning with
TensorFlow” Third International Conference on Computing and
Network Communications, 2020 Elsevier.
[7] Cheng, K.P., Mohan, R.E., Nhan, N.H.K., Le, A.V.: Multiobjective
genetic algorithm-based autonomous path planning for hinged-tetro
reconfigurable tiling robot. IEEE Access 8, 121267 (2020).
[8] Abdolrasol, M.G.M., Hussain, S.M.S., Ustun, T.S., Sarker, M.R.,
Hannan, M.A., Mohamed, R., Ali, J.A., Mekhilef, S., Milad, A.
“Artificial Neural Networks Based Optimization Techniques: A
Review”. Electronics 2021, 10, 2689, Elsevier.

Authorized licensed use limited to: College of Engineering - THIRUVANANTHAPURAM. Downloaded on August 17,2023 at 08:53:18 UTC from IEEE Xplore. Restrictions apply.