Abstract

This work presents a novel algorithm for local path planning for autonomous vehicles (AVs) which prioritizes both safety and adherence to traffic regulations, addressing critical functions for AV navigation, such as navigating complex environments, avoiding obstacles, and ensuring passenger and road users safety. The algorithm integrates the Twin Delayed Deep Deterministic Policy Gradient (TD3) with sensor fusion based on Nvidia Convolutional Neural Network (NCNN). The study utilizes the CARLA simulator, and real-world datasets, including KITTI and WAYMO, to train and evaluate the proposed algorithm. The proposed algorithm leverages the complementary strengths of Imitation Learning (IL) and Deep Reinforcement Learning (DRL) techniques, IL utilizes human driving data to provide the DRL agent with a foundation for safe and rule-abiding behavior, while the DRL agent refines its decision-making capabilities through real-time interaction with the environment. This combined approach aims to overcome limitations associated with individual techniques, such as long training time for DRL and lack of generalizability of supervised learning methods. Results from the CARLA simulations demonstrated the effectiveness of the proposed method and sensor fusion in obstacle detection and navigation precision. Realworld testing further validated the model’s ability to generalize from simulated environments to actual driving conditions, highlighting its potential for practical deployment in autonomous vehicles.

School

School of Sciences and Engineering

Department

Mechanical Engineering Department

Degree Name

MS in Robotics, Control and Smart Systems

Graduation Date

Summer 6-15-2024

Submission Date

7-17-2024

First Advisor

Amr El-Mougy

Committee Member 1

Sherif Aly

Committee Member 2

Mervat Abo El Kheir

Committee Member 3

Karim Seddik

Extent

125 p.

Document Type

Master's Thesis

Institutional Review Board (IRB) Approval

Not necessary for this item

Download

Share

COinS