SmartSort

AI-Powered Autonomous Waste Segregation Robot

YOLOv8 Vision · Robotic Arm · IoT Smart Bin Dashboard

Final Year Engineering Major Project Report

Mechatronics / Robotics | Team of 4 | Budget ■15,000

Parameter Value

■ Domain ■ Smart City / Urban

■ Team Size 4 Members

■ Budget ■9,760 (within ■15,000)

■ Duration 8 Months

■ Complexity Medium-Hard

■ Goal Major Project + Research Paper

1. PROBLEM STATEMENT

"Existing waste management in India relies on manual unsegregated collection — creating

environmental hazards, health risks, and massive recycling losses. There is no affordable,
AI-driven, autonomous segregation system accessible at the institutional or municipal level."

1.1 The Real-World Context

India generates over 62 million tonnes of solid waste annually. Less than 20% is processed scientifically. The core
problem is not collection — it is segregation. Mixed waste at the source makes downstream recycling nearly
impossible and contaminates landfills.
Current segregation methods fall into three inadequate categories:
• Manual sorting — workers hand-separate waste, exposing them to toxic, infectious and hazardous material
daily. Error-prone and unscalable.
• Sensor-based dumb bins — use inductive/IR sensors to detect metal vs non-metal only. Cannot distinguish
plastic from paper, or glass from e-waste.
• Industrial sorting lines — cost ■50 lakh to ■5 crore. Require dedicated infrastructure. Completely
inaccessible to colleges, hospitals, or small municipalities.

1.2 The Core Research Gap

Gap Description Impact

Visual AI Absent Sensor-based systems classify only 2–3 waste types (metal/non-metal)
60–70% waste misclassified

No Autonomy Robots require constant human teleoperation Labour cost remains high

No IoT Monitoring No bin fill-level tracking or alert systems Overflow goes unnoticed

High Cost Industrial systems cost ■50L+ Inaccessible to 99% of institutions

No 6-Class System No low-cost system handles plastic/metal/paper/glass/organic/e-waste

Recycling chains broken

1.3 Formal Problem Statement

"Existing waste segregation systems either rely on rudimentary proximity sensors with poor
classification granularity, or are industrial-scale solutions priced beyond institutional reach. There
is a critical need for a low-cost (sub-■15,000), autonomous, AI-enabled robotic system capable of
visually classifying waste into 6 categories in real-time using computer vision, physically sorting
items via a robotic arm, and monitoring bin fill levels through an IoT dashboard — all without
human intervention or cloud connectivity."
2. PROPOSED SOLUTION — SmartSort

SmartSort is a tabletop autonomous waste segregation system consisting of three integrated subsystems working
in a continuous pipeline:

Step Action Technology

1 Waste item placed on conveyor belt; belt moves item to detection zone DC Gear Motor + L298N

2 Overhead camera captures real-time image of item USB 1080p Camera

3 YOLOv8-nano AI model classifies item into 1 of 6 waste categories YOLOv8-nano on Raspberry Pi 5

4 3-DOF servo arm picks item and places into correct labelled bin MG996R Servos + PCA9685

5 Ultrasonic sensors monitor fill level of each bin continuously HC-SR04 × 6

6 Flask IoT dashboard shows live feed, classification log, bin levels Flask + WiFi

7 Telegram alert sent when any bin exceeds 80% fill level Telegram Bot API

2.1 The 6 Waste Categories

Class Label Examples Bin Colour Real-World Destination

plastic Bottles, bags, cups, packaging Blue Recycling plant

metal Cans, foil, tins, bottle caps Grey Metal recycler

paper Newspaper, cardboard, boxes White Paper recycler

wet_organic Food scraps, peels, tea bags Green Compost / biogas

glass Bottles, jars, broken glassware Red Glass recycler

e_waste Batteries, wires, PCBs, phone parts Black E-waste facility

★ KEY INNOVATION: Achieving clean 6-class AI-based visual waste separation with a physical
robotic arm and live IoT monitoring — all under ■10,000 hardware cost — is not achieved by any
existing published low-cost student-level project.
3. SYSTEM ARCHITECTURE

SMARTSORT SYSTEM ARCHITECTURE [USB Camera 1080p] ■■■ [Raspberry Pi 5 (4GB)] [LED Strip x2]
■■■ | |■■ YOLOv8-nano (TFLite INT8) [HC-SR04 x6] ■■■ |■■ Bin Fill Monitor |■■ Flask IoT
Dashboard |■■ Telegram Alert Bot | [PCA9685 PWM Board] (I2C) |
■■■■■■■■■■■■■■■■■■■■■■ | 3-DOF Servo Arm | | Base + Shoulder + | | Elbow + Gripper |
■■■■■■■■■■■■■■■■■■■■■■ | ■■■■■■■■■■■▼■■■■■■■■■■■ | 6 Segregation Bins | |
(Plastic/Metal/Paper | | /Glass/Organic/ | | E-Waste) | ■■■■■■■■■■■■■■■■■■■■■■■
[L298N + DC Motor] ■■■ Conveyor Belt Power: 5V 5A (Pi) + 6V 3A BEC (Servos) — separated
supplies

3.1 Complete Technology Stack

Layer Technology Purpose

Compute Raspberry Pi 5 (4GB) Runs vision, arm control, dashboard

Vision USB Camera 1080p (overhead) Captures waste in detection zone

AI Model YOLOv8-nano → TFLite INT8 Real-time classification at 15+ FPS

Arm Control MG996R × 4 + PCA9685 Driver Pick and place waste into correct bin

Conveyor DC Gear Motor + L298N Moves waste to detection zone

Bin Sensing HC-SR04 Ultrasonic × 6 Measures fill level per bin

IoT Dashboard Flask + WiFi (Pi onboard) Live feed, log, bin levels on browser

Alerts Telegram Bot API (free) Phone alert when bin >80% full

Training Google Colab T4 GPU (free) Fine-tune YOLOv8-nano on dataset

Dataset TrashNet + TACO + custom 6-class waste images (5,000+)

Chassis 3D-printed PLA + TPU gripper Frame, arm mount, bin housing

Power 5V 5A USB-C + 6V 3A BEC Separated Pi and servo power

4. TEAM DIVISION OF WORK

Member Role Responsibilities

Chassis design (Fusion 360), 3D printing conveyor frame,

Member 1 Mechanical Lead
bin housing, arm mount, waterproofing

Servo wiring, PCA9685 setup, L298N motor driver,

Member 2 Electronics Lead
HC-SR04 sensors, power circuit, BEC configuration

Dataset collection, YOLOv8 training on Colab,

Member 3 AI / Vision Lead
TFLite INT8 export, live inference on Pi 5

Arm kinematics (Python), Flask dashboard,

Member 4 Software / IoT Lead
Telegram bot, conveyor control logic, SQLite logging

All 4 members participate equally in integration testing and final demo (Month 6–8).

5. DATASET STRATEGY

No large custom dataset is required. The following three-source strategy ensures high accuracy without months of
data collection:

Source Images Classes Action

TrashNet (Stanford, free GitHub) 2,527 6 (glass/paper/cardboard/plastic/metal/trash)

Primary backbone — download and use directly

TACO (free, annotated) 1,500+ 60 categories → map to 6 Improves real-world litter accuracy

Custom Collection (your campus) 300–500 6 Capture under your exact camera + lighting for demo

Roboflow Augmentation (free tier)~5,000+ 6 Auto-augment: flip, brightness, crop, noise — multiply

Training: Google Colab T4 GPU (free). Fine-tune YOLOv8-nano from COCO pretrained weights. Estimated
training time: 1.5–2 hours. Target: 88–92% mAP@0.5 on test set.
6. MONTH-BY-MONTH IMPLEMENTATION PLAN

Month Phase Key Deliverable

1 Design + Component Order Finalize system design; order all components; setup Raspberry Pi 5 OS + YOLOv8 environmen

2 Mechanical Build 3D print conveyor frame, arm mount, 6-bin housing; assemble conveyor belt; test servo arm ma

3 AI Model Development Download TrashNet + TACO; augment on Roboflow; train YOLOv8-nano; verify 15+ FPS on P

4 Arm Control + Kinematics Implement 3-DOF forward kinematics; map 6 bin positions to servo angles; test pick-place sequ

5 Sensor Integration + IoT Wire HC-SR04 × 6; build Flask dashboard with live feed + bin fill bars; configure Telegram bot

6 Full Integration Testing End-to-end pipeline with real mixed waste; measure accuracy, pick success rate, cycle time

7 Refinement + Documentation Fix edge cases; improve gripper; write IEEE-format report; record 5-min demo video

8 Final Demo + Submission Live demo for evaluators; cost analysis + scalability pitch; submit to national conference

7. EXPECTED CHALLENGES + SOLUTIONS

Challenge Root Cause Solution

Arm misses small items Servo positioning inaccuracy Optical end-stop sensors for homing; PID position control

Camera glare / uneven shadows

Ambient lighting variation 2 diffused LED strips on sides of detection zone

Wet organic waste sticks to gripper

Gripper absorbs moisture TPU (flexible) 3D-printed gripper + silicone spray coating

Model confuses plastic bag vsSimilar

paper colour/texture in training images
Add 100+ bag-specific images; use bounding box area as secondary feature

Servo jitter under load Insufficient current from Pi GPIO Dedicated 6V 3A BEC for servo power — never power from Pi pins

Bin sensor false readings Ultrasonic reflections from angled Angle

waste sensor 15° downward; 5-reading median filter in software
8. COMPLETE BILL OF MATERIALS (BOM)

Component Qty Unit Price (INR) Total (INR)

Raspberry Pi 5 (4GB) 1 5,200 5,200

USB Camera 1080p 1 700 700

MG996R Servo Motors 4 200 800

PCA9685 Servo Driver Board 1 220 220

L298N Motor Driver Module 1 80 80

DC Gear Motor (conveyor) 1 120 120

HC-SR04 Ultrasonic Sensors 6 40 240

PLA Filament 1 kg 1 700 700

TPU Filament 250 g (gripper) 1 350 350

Rubber conveyor belt strip 1 150 150

5V 5A USB-C Power Supply (Pi) 1 400 400

6V 3A BEC for Servos 1 180 180

LED Strips x 2 (detection zone) 1 120 120

Jumper wires, PCB, connectors - 300 300

Misc (nuts, bolts, standoffs) - 200 200

TOTAL HARDWARE COST ■9,760

Budget Buffer: ■9,760 spent of ■15,000 available. ■5,240 reserved for part replacements, reprint
iterations, and contingency.
9. INNOVATION FACTOR — What Makes SmartSort Different

Feature Existing Projects SmartSort

Detection method IR + inductive sensors YOLOv8 camera vision AI

Waste categories 2–3 (metal / non-metal) 6 classes (full spectrum)

Works on mixed waste? No Yes

Bin fill monitoring No Yes — live IoT dashboard

Mobile alerts No Yes — Telegram bot

Segregation accuracy ~60–70% (sensor-based) Target 88–92% (AI-based)

Cost ■5K–■8K (basic sensor bot) ■9,760 (full AI system)

Offline operation Usually cloud-dependent 100% offline — no internet needed

9.1 Five Core Innovation Points

• 6-class visual AI classification — No published low-cost student project achieves clean 6-class
YOLOv8-based separation with a physical sorting arm.
• Full mechatronics integration — Mechanical design + electronics + computer vision + IoT in a single
coherent system.
• Completely demoable on any table — No pipes, no outdoor access, no special environment. Works in the
lab on Day 1 of evaluation.
• Offline-first architecture — Flask dashboard runs on local WiFi; Telegram bot is the only cloud dependency
and is fully optional.
• Publication-ready — Directly addresses a documented gap in IEEE/MDPI literature on AI-based waste
management systems.

10. FUTURE SCOPE & SCALABILITY

10.1 Short-Term Extensions (within project timeline)

• Add weight sensor (load cell) per bin for precise fill-level monitoring in grams.
• Voice announcement of classification result using edge TTS (speaks 'Plastic detected').
• Extend to 8 categories: add 'hazardous' (syringes, chemicals) and 'textile' classes.

10.2 Medium-Term (Post-graduation / startup)

• Replace single-arm with dual-arm setup — doubles throughput for canteen/cafeteria deployment.
• Add RFID/barcode reader — scan packaged waste to supplement AI classification with product data.
• Municipal-scale version: conveyor line with multiple sort arms — same AI, larger form factor.

10.3 Market Opportunity

The AI in Waste Management and Recycling Market was valued at $3.5 billion in 2023 and is forecast to reach
$15.9 billion by 2031. SmartSort's architecture is the seed of a viable deep-tech startup targeting Smart City
missions, Swachh Bharat infrastructure, and institutional waste management contracts.
10.4 Research Publication Pathway
• Submit dataset (custom ISL/waste images) as open-source — generates citations independently.
• Target conferences: IEEE ICRA workshop, IETE National Conference, ISTE Symposium.
• Target journals: MDPI Sensors, IEEE Access (open-access, no publication fee for students).
11. WHY THIS PROJECT WILL IMPRESS EVALUATORS

For Your Professors

• Full mechatronics integration — mechanical + electronics + AI + IoT in one project.
• Live demo works right on the lab table; no setup, no special environment needed.
• Backed by published IEEE/MDPI research with clear citation trail.
• Structured IEEE-format report maps every design choice back to the problem statement.

For Industry Recruiters

• Demonstrates computer vision, embedded AI, IoT, and robotics control — all top hiring skills in 2025.
• Scalable concept clearly pitchable: ■10K lab demo → ■10 crore smart city contract.
• Real AI deployment (not simulation) — YOLOv8 running live at 15 FPS on edge hardware.

For Hackathons (SIH, Ideathon, etc.)

• Directly maps to Swachh Bharat Mission problem themes.
• Maps to MeitY Smart City and Urban Local Body (ULB) digital infrastructure tracks.
• Under ■10K hardware cost — instantly credible as a scalable, low-cost innovation.

BOTTOM LINE: SmartSort gives your team of 4 a project where every member has a clear,
meaningful role — Mechanical, Electronics, AI, and Software — that naturally converges into a
single impressive, live-demoable system. You are not building a toy. You are building a real
solution to a real problem, with a real AI model, at a price point that makes it deployable.

Report generated for Final Year Engineering Major Project. All cost estimates in INR as of 2025. Component prices may
vary by ±10% by vendor.