README - Python Client for TORCS Telemetry

Overview

This project is a Python client for TORCS (The Open Racing Car Simulator), designed to collect telemetry data from the racing simulation. The collected data is stored in a CSV file for further AI-based controller training.

Features:

• Real-time telemetry data collection from TORCS.
• Storage of relevant vehicle and track data for analysis.
• CSV logging for AI-based training and performance evaluation.
• Noisy sensors (focus, opponent, and track) are disabled for clean data collection.

Stored Data in CSV

The script stores relevant telemetry data from the TORCS simulation for AI model training. Below is a breakdown of what is stored and what is ignored.

Stored Data:

1. Core Vehicle State (Single-Value Sensors)

Field	Description
curLapTime	Current lap time in seconds.
distFromStart	Distance from the start line (meters).
distRaced	Total distance raced (meters).
fuel	Remaining fuel (irrelevant as fuel is unlimited).
gear	Current gear in use.
lastLapTime	Time taken for the last completed lap.
racePos	Current position in the race.
rpm	Engine revolutions per minute (RPM).
speedX	Forward speed of the car.
speedY	Sideways speed.
speedZ	Vertical speed.
trackPos	Position relative to the center of the track.
angle	Angle between the car's heading and the track axis.
z	Car's altitude above the track.

2. Wheel Spin Velocity (4 Values)

Field Format	Description
wheelSpin_0-3	Measures wheel slip/spin for each wheel.

• Useful for detecting oversteering or traction loss.

Data NOT Stored:

1. Noisy Sensors (disabled as per requirements):

   • focus (Focus Sensors)
   • opponent (Opponent Sensors)
   • track (Edge-Distance Sensors)

2. Car Damage

   • The damage field is ignored as car damage is disabled.

3. Fuel Level

   • The fuel field is ignored since the project requirements state that fuel is unlimited.

Telemetry Implementation

The telemetry data collection is implemented in a file called telemetry.py, which is responsible for the storage of the telemetry data recieved from the TORCS simulator. This file is imported and executed by msgParser, which processes and formats the received data before logging it.

License

This project is part of the Artificial Intelligence AI-2002 Spring-2025 course and is intended for academic purposes only.

Installation

1. Clone the repository:

   git clone https://github.com/hafsa-imtiaz/torcs-python-client

2. Run the TORCS Simulator in the background.
3. Run the pyClient.py file.

Original Code Source

The original Python 2 code for pyClient is from: http://games.ws.dei.polimi.it/competitions/scr/ and has been converted to Python 3 with telemetry updates in telemetry.py.

---

New Files and AI Controller Integration

This project has been extended with an AI-based autonomous driving controller trained using supervised learning. The following new files have been added:

1. CarANN.py

Purpose:
Loads the trained Artificial Neural Network (ANN) model and RobustScaler, and predicts real-time control outputs using live telemetry input.
Role in Project:
Serves as the AI driver, replacing manual or rule-based control logic.

2. model_training.ipynb

Purpose:
Contains the full training pipeline for the AI controller, including preprocessing, feature engineering, model architecture, training, and evaluation.
Role in Project:
Used to train the model on recorded telemetry data with a hybrid LSTM-Dense architecture.

3. racing_scaler.pkl

Purpose:
Stores the fitted RobustScaler used for normalizing input features during training and inference.
Role in Project:
Ensures consistent feature scaling across both training and real-time prediction.

4. torcs_model.h5 and torcs_model.tflite

Purpose:
Serialized versions of the trained neural network.
Role in Project:
Used by CarANN.py to make predictions on sensor input.

Note: An expanded telemetry dataset (~350,000 frames) including 72 input features and 5 control outputs has been used. The report contains the complete model and training details.