Wall-Following Robot Navigation Classification

A machine learning project for classifying movement commands of a wall-following mobile robot (SCITOS-G5) based on ultrasound sensor readings.

📋 Table of Contents

• Project Overview
• Dataset Description
• Data Insights
• Installation
• Project Structure
• Usage
• Methodology
• Experimental Results
• Key Insights
• References
• License

🤖 Project Overview

This project aims to classify the movement commands of a wall-following mobile robot (SCITOS-G5) based on ultrasound sensor readings. The robot navigates a room following walls in a clockwise direction, and the task is to predict its movement at each time step:

• Move-Forward
• Slight-Right-Turn
• Sharp-Right-Turn
• Slight-Left-Turn

The classification problem is non-linearly separable, making it an interesting benchmark for supervised machine learning algorithms.

📊 Dataset Description

The dataset comes from the UCI Machine Learning Repository and contains three versions based on the number of input sensors:

1. Full Sensor Readings (sensor_readings_24.csv)

• 24 ultrasound sensors arranged circularly around the robot's "waist"
• Each sensor: US1 – US24 with a reference angle from front (180°) clockwise
• Samples: 5,456
• Attributes: 24 numeric features + 1 class label

2. Simplified 4-Sensor Readings (sensor_readings_4.csv)

• Sensors aggregated into arcs: SD_front, SD_left, SD_right, SD_back
• Each represents minimum distance in 60° arc
• Samples: 5,456
• Attributes: 4 numeric features + 1 class label

3. Simplified 2-Sensor Readings (sensor_readings_2.csv)

• Only SD_front and SD_left
• Samples: 5,456
• Attributes: 2 numeric features + 1 class label

Class Distribution

Class	Samples	Percentage
Move-Forward	2,205	40.41%
Slight-Right-Turn	826	15.13%
Sharp-Right-Turn	2,097	38.43%
Slight-Left-Turn	328	6.01%

Missing values: None

📈 Data Insights

Summary Statistics - 24-Sensor Dataset

Sensor	Max	Min	Mean	SD
US1	5.0000	0.40000	1.47162	0.80280
US2	5.0250	0.43700	2.32704	1.41015
US3	5.0290	0.47000	2.48935	1.24743
US4	5.0170	0.83300	2.79650	1.30937
US5	5.0000	1.12000	2.95855	1.33922
US6	5.0050	1.11400	2.89307	1.28258
US7	5.0080	1.12200	3.35111	1.41369
US8	5.0870	0.85900	2.54040	1.11155
US9	5.0000	0.83600	3.12562	1.35697
US10	5.0220	0.81000	2.83239	1.30784
US11	5.0190	0.78300	2.54940	1.38203
US12	5.0000	0.77800	2.07778	1.24930
US13	5.0030	0.77000	2.12578	1.40717
US14	5.0000	0.75600	2.19049	1.57687
US15	5.0000	0.49500	2.20577	1.71543
US16	5.0000	0.42400	1.20211	1.09857
US17	5.0000	0.37300	0.98983	0.94207
US18	5.0000	0.35400	0.91027	0.88953
US19	5.0000	0.34000	1.05811	1.14463
US20	5.0000	0.35500	1.07632	1.14150
US21	5.0000	0.38000	1.01592	0.88744
US22	5.0000	0.37000	1.77803	1.57169
US23	5.0000	0.36700	1.55505	1.29145
US24	5.0000	0.37700	1.57851	1.15048

Summary Statistics - 4-Sensor Dataset

Feature	Max	Min	Mean	SD
SD_front	5	0.49500	1.29031	0.62670
SD_left	5	0.34000	0.68127	0.34259
SD_right	5	0.83600	1.88182	0.56253
SD_back	5	0.36700	1.27369	0.82175

Summary Statistics - 2-Sensor Dataset

Feature	Max	Min	Mean	SD
SD_front	5	0.49500	1.29031	0.62670
SD_left	5	0.34000	0.68127	0.34259

⚠️ Note: These statistics show sensor ranges and variability, which are the features fed into the machine learning models.

🛠️ Installation

Prerequisites

• Python 3.7 or higher
• pip package manager

Setup

1. Clone the repository:

git clone https://github.com/bugsNburgers/wall-following-robot.git
cd wall-following-robot

2. Install required dependencies:

pip install -r requirements.txt

Required Packages

• numpy
• pandas
• scikit-learn
• matplotlib
• seaborn
• joblib

📁 Project Structure

wall-following-robot/
├── data/                           # Dataset files
│   ├── sensor_readings_2.csv       # 2-sensor dataset
│   ├── sensor_readings_4.csv       # 4-sensor dataset
│   └── sensor_readings_24.csv      # 24-sensor dataset
├── src/                            # Source code
│   ├── loader.py                   # Data loading utilities
│   ├── train_and_eval.py          # Model training and evaluation
│   └── demo.py                     # Demo script for predictions
├── models/                         # Trained model files (generated)
├── outputs/                        # Evaluation outputs (generated)
│   ├── summary_metrics_*.csv      # Performance metrics
│   └── confusion_matrix_*.png     # Confusion matrices
├── requirements.txt                # Python dependencies
└── README.md                       # This file

🚀 Usage

Training Models

To train all models on all datasets:

python -m src.train_and_eval

This will:

• Train SVM, Random Forest, and Logistic Regression models
• Evaluate on test sets
• Generate confusion matrices
• Save models to models/ directory
• Save performance metrics to outputs/ directory

Making Predictions

To make predictions on specific data points:

# Predict using a specific model on the full dataset
python -m src.demo data/sensor_readings_4.csv sensor_readings_4_svm_rbf 0

# Predict using a specific model on the test set
python -m src.demo data/sensor_readings_4.csv sensor_readings_4_random_forest 10 test

Arguments:

• <path_to_csv>: Path to the dataset file
• <model_name>: Model filename (without .joblib extension)
• <row_index>: Index of the row to predict
• test (optional): Use test set instead of full dataset

Example Workflow

# 1. Train models on all datasets
python -m src.train_and_eval

# 2. View results
cat outputs/summary_metrics_sensor_readings_4.csv

# 3. Test prediction with trained model
python -m src.demo data/sensor_readings_4.csv sensor_readings_4_random_forest 0 test

🔬 Methodology

Preprocessing

• Data split: Train-Test Split (80%-20%)
• Encoding classes: Class labels mapped to integers:
  • Move-Forward → 0
  • Slight-Right-Turn → 1
  • Sharp-Right-Turn → 2
  • Slight-Left-Turn → 3
• Feature scaling: StandardScaler applied for SVM and Logistic Regression; not required for Random Forest

Models Implemented

1. Support Vector Machine (SVM)

   • Kernel: RBF (Radial Basis Function)
   • Class weights: Balanced

2. Random Forest Classifier

   • Number of estimators: 200
   • Class weights: Balanced

3. Logistic Regression

   • Solver: lbfgs
   • Multi-class: auto
   • Max iterations: 500

Evaluation Metrics

• Accuracy: Overall correctness
• Precision: Correct positive predictions / Total predicted positives
• Recall: Correct positive predictions / Total actual positives
• F1-Score: Harmonic mean of precision and recall

📊 Experimental Results

Dataset	Model	Accuracy	Precision	Recall	F1
2-Sensor	SVM (RBF)	0.9588	0.9364	0.9715	0.9520
2-Sensor	Random Forest	1.0000	1.0000	1.0000	1.0000
2-Sensor	Logistic Regression	0.9167	0.9046	0.9298	0.9122
4-Sensor	SVM (RBF)	0.9469	0.9306	0.9641	0.9450
4-Sensor	Random Forest	1.0000	1.0000	1.0000	1.0000
4-Sensor	Logistic Regression	0.9112	0.8806	0.9233	0.8968
24-Sensor	SVM (RBF)	0.8773	0.8423	0.9016	0.8662
24-Sensor	Random Forest	0.9927	0.9902	0.9839	0.9870
24-Sensor	Logistic Regression	0.6850	0.6422	0.7521	0.6676

Observations

As the number of sensors increases, Logistic Regression struggles due to non-linear separability and more complex feature interactions. Random Forest consistently achieves near-perfect accuracy. SVM performs well but slightly drops with 24 sensors.

💡 Key Insights

1. Random Forest is robust to input feature size and non-linearity, achieving near-perfect classification across all datasets.

2. SVM works well on smaller sensor sets but slightly degrades on full 24-sensor data, though still maintaining good performance.

3. Logistic Regression performs decently on simplified datasets but struggles on the full 24-sensor data, confirming the non-linear nature of the task.

4. Sensor statistics explain difficulty: Increased variance and overlapping sensor ranges make linear separability impossible as the number of sensors grows.

5. Simpler can be better: The 2-sensor and 4-sensor datasets achieve excellent performance while being computationally more efficient.

📚 References

• Lichman, M. (2013). UCI Machine Learning Repository. http://archive.ics.uci.edu/ml
• Freire, A. L., Barreto, G. A., Veloso, M., & Varela, A. T. (2009). Short-Term Memory Mechanisms in Neural Network Learning of Robot Navigation Tasks: A Case Study. LARS'2009, Valparaíso, Chile. DOI: 10.1109/LARS.2009.5418323

📄 License

Dataset: UCI Machine Learning Repository – "Other" license

---

This project demonstrates the application of machine learning algorithms to robotics navigation problems and serves as a benchmark for classification algorithms on non-linearly separable data.