Autonomous Vehicle Steering Control via End-to-End Deep Learning in BeamNG.tech

Overview

This project explores end-to-end deep learning for autonomous steering using the BeamNG.tech simulator. The goal is to predict steering angles directly from camera images in real-time.

Features

• Predicts steering angles from front and side camera images
• Lane recovery using side camera offsets
• Three CNN architectures tested, including residual and attention mechanisms
• Data augmentation for more robust learning

Models

• Implemented and trained 3 models each with different number of layers and structure
• Model 3 was the best I achieved which predicted angles better than the REST (but still not trust worthy_

Training Details

• Loss Function: Mean Squared Error (MSE)
• Optimizer: Adam (LR=0.002) with weight decay
• Data Augmentation: Random flip, translation, brightness changes
• Validation: 25% of dataset reserved
• Note: Models were partially trained due to limited computation power, which affected performance

Results

• Note: Car was purposely parked at an angle to check if it goes into lane

Training Phase:

Model 1 Demo: (Baseline so it tries to maneuver towards the left but fails and crashes while steering right)

Recording.2025-12-18.222157.mp4

Model 2 Demo: (Performed Slightly better than Model 1, i.e is able to steer its way towards the road trying to avoid the obstacles)

Video.Project.2.mp4

Model 3 Demo: (Performed better than the previous ones. Although not perfect but still managed to go in a straight line from an angled parking postion)