Maters Thesis Presentation - Vlad Muscoi

Title: Autonomous Vehicle Navigation in a Digital Twin City Using Reinforcement Learning and Multi-Modal Perception in Unreal Engine

Abstract:

Developing safe and reliable autonomous vehicles requires extensive testing across diverse and complex traffic scenarios, a process that is costly, time-consuming, and potentially dangerous when conducted in the real world. This thesis explores the use of Unreal Engine 5 (UE5) as a high-fidelity simulation platform for training and evaluating an autonomous driving pipeline that combines multi-modal perception with reinforcement learning.

The proposed system consists of two main components. The first is a perception module based on BEVFusion, a model that fuses six RGB camera images and one LiDAR point cloud into a Bird's-Eye View (BEV) segmentation map: a compact, top-down representation of the vehicle's surroundings. To train this model, a synthetic dataset was collected using UE5's City Sample project, which leverages the engine's latest rendering capabilities and built-in traffic simulation to produce visually realistic and semantically rich driving scenes. The model was retrained on six semantic classes, achieving performance comparable to other models trained on real-world or synthetic data.

The second component is a reinforcement learning agent trained using Proximal Policy Optimization. Operating within a T-intersection environment, the agent learns to execute a complete driving maneuver: approaching the intersection, making a right turn, maintaining a target speed, and avoiding collisions with both parked and moving vehicles while staying within legal lane boundaries. The agent successfully learns to complete this task reliably, demonstrating the viability of UE5 as an RL training environment.

This work demonstrates that Unreal Engine 5 is a capable platform for developing and validating autonomous driving systems, offering a scalable and cost-effective alternative to real-world data collection and testing.

Supervisors: Jirí Kosinka, Kailai Li