Autonomous Driving in a Simulated Racetrack

Overview

This project demonstrates the development and implementation of an autonomous driving system in a simulated racetrack environment. The system leverages EfficientNet-B0 as a backbone for lane detection and a vision-based controller for end-to-end driving, optimized within the IGN Gazebo simulation environment.

---

Project Details

Autonomous Driving System

• Backbone: EfficientNet-B0
• Input Data: RGB images from a front-facing camera
• Environment: Sonoma Raceway simulated in IGN Gazebo
• Pipeline:
  • RGB camera images are fed into the EfficientNet-B0 backbone for lane detection and trajectory estimation.
  • The vision-based controller dynamically adjusts steering and velocity for end-to-end autonomous driving.
  • Commands are executed in real-time at 10 Hz for smooth driving.

Training:

• Pre-processed racetrack images using standard augmentation techniques to enhance generalization.
• The EfficientNet-B0 model was fine-tuned for racetrack-specific data using transfer learning on the TuSimple dataset.
• The controller was optimized for robust trajectory tracking using closed-loop simulations.

---

Simulation Environment

• Platform: IGN Gazebo
• Physics Engine: ODE for realistic vehicle dynamics.
• Track: Sonoma Raceway, a complex circuit featuring varying curvature and elevation.

Tools and Frameworks

• Deep Learning Framework: PyTorch
• Robot Operating System: ROS2 (Humble)
• Simulation: IGN Gazebo

---

Key Features

1. EfficientNet-B0 Backbone for Autonomous Driving:

   • Lightweight and efficient deep learning model.
   • Processes RGB camera images in real-time to detect lanes and estimate trajectories.

2. Vision-Based Controller:

   • Executes steering and velocity commands based on model predictions.
   • Designed for robust autonomous driving at 10 Hz, ensuring smooth operation in complex racetrack scenarios.

3. IGN Gazebo Integration:

   • High-fidelity simulation with accurate racetrack geometry.
   • Seamless integration with ROS2 for data flow and control commands.

---

Results

• Demonstrated fully autonomous driving in the Sonoma Raceway environment with accurate trajectory tracking.
• The system processes camera images and executes driving commands at 10 Hz, enabling real-time autonomous performance.
• Achieved robust driving across varying lighting and weather conditions in the simulated racetrack.
• Successfully handled the dynamic complexities of a high-curvature racetrack, maintaining stability and consistent performance.