Jewel Nath
← Home

EnviroBot

A rover with a six-servo arm that finds waste with a camera and picks it up off the floor. Built at Comilla University with my team: a YOLOv8 model labels plastic, paper and trash, and an Arduino drives the arm and the wheels.

Built2023 – 2024My part: with the team, its design, build and tests; I also put its ROS 2 model online and wrote its report

What it shows

Waste detection
A YOLOv8 model trained on a 1,700-image dataset in three classes; in the team’s evaluation it labelled 91–94% of each class correctly
Mobile manipulation
A six-servo arm on a steel rover that reaches down and picks objects up off the floor
Two controllers
In its design, vision on a Raspberry Pi and all motion on an Arduino, joined by a serial link, with a phone app over Bluetooth
ROS 2
The arm modelled in URDF and moved joint by joint in RViz

Built · Sep 2023 – Jan 2024

A rover with an arm

A four-wheel rover on a stainless steel base carries a six-servo aluminium arm with a gripper, a mesh bin and a camera. The arm sits at the front, so it reaches the floor in front of the robot without the base in its way.

Picking a cube up off the floor · Jan 2024

In video

Short silent clips. Tap one to play it.

The detector and the robot

  • Jan 2024

    The detector running live on a laptop, boxing and naming objects as they come into view. This version also has a metal class.

  • Jan 2024

    An arm test on the floor: the arm reaches down to a small box. The rover is on a power cable.

  • Jan 2024

    The finished robot up close: the arm, the bin, the battery and the boards

Later

  • May 2024

    A driving test of the rover with its arm, on mecanum wheels

The build

  1. By Jan 2024

    The steel base with a gear motor at each corner, from below and above.

  2. By Jan 2024

    The arm fitted at the front of the base, before the wiring.

  3. Jan 2024

    Finished, from the front: the camera, the bin and the boards.

EnviroBot from the earliest dated image of the arm’s model to the later driving test (from the report)

How it works

A camera, two boards and six servos

The robot splits its work between two boards. In its design, a Raspberry Pi reads the camera and runs the detector; an Arduino Uno moves the arm and the wheels, and takes commands from the Pi over a serial link and from a phone over Bluetooth.

Datasheet

Base
A 14-inch stainless steel plate, a 12 V gear motor at each corner
Arm
An aluminium six-servo kit: waist, shoulder, elbow, wrist pitch and roll, gripper
Vision
A USB webcam under the arm, and a Raspberry Pi 3 B in the design
Motion
An Arduino Uno, a PCA9685 servo driver and two BTS7960 motor drivers
Control
A phone app made in MIT App Inventor, over Bluetooth
Power
A 3-cell LiPo battery, 11.1 V, 2200 mAh
Software
YOLOv8 trained in Google Colab; the arm in URDF for ROS 2 Humble

The whole robot

System overview, from the EnviroBot report. Tap to open it large.

How well it finds waste

Team’s evaluationIn the team’s evaluation of the detector: the share of each class it labelled correctly.

of paper
91%
of plastic
94%
of trash
92%
images in its dataset, in three classes
1,700

Finding waste

The team’s dataset holds 1,700 images in three classes, plastic, paper and trash, many of them from an earlier project in the department. Each item is labelled with a box and a class; some images had their brightness changed or were shifted or rotated to enlarge the set, and all were resized to 512 by 384 pixels.

The detector is YOLOv8 in its large size, trained for 50 epochs in Google Colab. It finds the objects in an image and names them in one pass of the network.

The detector’s confusion matrix, from the team’s evaluation
From images to a trained detector
Training over 50 epochs

Driving it from a phone

Before any autonomy, the team made the robot fully controllable by hand. The app has a pair of buttons for each joint, a slider for the arm’s speed, and buttons to save a sequence of positions, run it back and reset; other buttons drive the rover.

Each motor and each servo was brought up on its own from the Arduino’s serial monitor. Every servo’s starting, lowest and highest angles were found and written down, so the arm’s commands stay within each joint’s range.

The app’s arm screen
Each command, from the phone to a joint or a wheel

Two boards

The Pi and the Arduino talk over a serial link. The Arduino reaches the servo driver over I²C and drives each pair of wheels through its own BTS7960. Keeping all motion on the Arduino meant the arm and the wheels could be driven from the phone without the Pi.

The wiring, pin by pin, from the team’s circuit diagram

The arm in ROS 2

The arm is modelled in URDF as a ROS 2 package: six controlled joints from the waist to the gripper, each with limits on its travel. The team first moved the model joint by joint in RViz, in September 2023.

The full autonomous pick-up was designed on top of it: the detector picks a target, MoveIt plans the arm’s path, and the Arduino moves the joints. The detector and the pick-up were each shown on video on their own; MoveIt stayed in the design, and joining them into one program is the next step.

The arm’s model in RViz, September 2023
The arm’s joints, their axes and their ranges
The autonomous pick-up, as designed

My part, and the team

Built at Comilla University

EnviroBot was built at Comilla University by Tasmiara Jahan Toma, Tanjina Akter and me, supervised by Amena Begum, Assistant Professor of ICT. We designed, built and tested it together; I put its ROS 2 model into its public repository and wrote its technical report.

The team

  • Jewel Nath
  • Tasmiara Jahan Toma
  • Tanjina Akter

The robot was presented at Comilla University in January 2024.

Built on

The robot’s ROS 2 model is in its public repository.

What I learned

Build the base for the arm

The arm, not the rover, set the design: we chose a steel base and four geared motors to carry it, at the cost of a heavier robot.

Manual control first

The phone app let us test and calibrate every joint and the drive before any autonomy.

Reuse data, and name the metric

An earlier project’s images, enlarged with augmentation, gave a usable detector. Recording which metric each number is, and keeping background images, would have made the evaluation stronger.

The report

Read the full design

14 pages and 12 diagrams, checked against the team's report, the circuit diagram, the repository and the dated photos and videos, with the team as co-authors.

Next: ConveySort-Arm

A conveyor belt and a six-servo arm that sort blocks by size, built at RoboAI in 2024, with the same arm model.