MORAL–Drive: Multi-Objective Reinforcement Learning for Autonomous Vehicle Longitudinal Control in Car-Following Scenarios

Overview

Autonomous vehicles must maintain safe and comfortable longitudinal control in car-following scenarios, balancing often competing objectives such as safety margins, ride comfort, and traffic efficiency. Classical car-following models rely on fixed rules or single-objective optimization, which struggle to adapt to the trade-offs drivers make in real traffic.

MORAL-Drive develops a multi-objective reinforcement learning framework for autonomous vehicle longitudinal control, explicitly optimizing for multiple, sometimes conflicting objectives simultaneously, and validating the resulting policies against real-world car-following behavior.

Objectives

The main objective of the MORAL–Drive project is the design, implementation, and experimental validation on real data of an intelligent longitudinal control system for autonomous vehicles based on multi-objective reinforcement learning (RL) algorithms that can respond efficiently, safely, and comfortably in car-following scenarios specific to urban traffic.

This is achieved through the following specific objectives:

1. Comparative analysis & conceptual frameworkComparative analysis of reinforcement learning architectures and development of a conceptual framework for autonomous vehicle longitudinal control based on multi-objective reinforcement learning, applicable to car-following scenarios.
2. Agent design & implementationDesign and implementation of a reinforcement learning agent guided by a multi-objective reward function that jointly optimizes passenger comfort, safety, and energy/fuel efficiency, implemented and tested in traffic simulation environments.
3. Validation against baselinesValidation of the RL agent against baseline car-following models on standard efficiency, safety, and comfort metrics, using real-world highway and urban traffic datasets.
4. Interpretability analysisAnalysis of the interpretability and decision-making behavior of the multi-objective RL agent using explainable AI methods, to increase transparency and understanding of control decisions.
5. DisseminationDissemination of research results through ISI Web of Science and BDI-indexed articles, scientific presentations, and reporting in line with AOSR requirements.

Expected Outputs

Based on the correlation between the project objectives and the proposed activities, the publication of a minimum of seven scientific articles is estimated, according to the following performance indicators (KPIs):

1. Four conference papersPresented at scientific events organized by the Academy of Romanian Scientists in 2026 and 2027.
2. Two journal articlesAccepted in impact-factor journals, corresponding to project deliverables D2.1 and D4.4.
3. One journal article (IF ≥ 2.0, Web of Science)Manuscript submitted with confirmed receipt by the journal editorial office by December 4, 2027, to be included as evidence in the final report (corresponds to deliverable D5.3).

Progress & results

Project team

Mădălin-Dorin Pop

Principal Investigator

Department of Computer and Information Technology
Politehnica University Timișoara

madalin.pop[at]upt.ro Google Scholar profile

PC

Petra Csereoka

Researcher

Department of Computer and Information Technology
Politehnica University Timișoara

petra.csereoka[at]upt.ro Google Scholar profile

MK

Mátyás Kuti-Kreszács

Researcher

Faculty of Mathematics and Computer Science
Babeş-Bolyai University, Cluj-Napoca

matyas.kuti[at]ubbcluj.ro Google Scholar profile

Publications from this project

Funding

This research project is funded by the Academy of Romanian Scientists through the AOSR-TEAMS-IV Young Researchers Research Projects Competition (2026–2027 Edition), Project “MORAL-Drive: Multi-Objective Reinforcement Learning for Autonomous Vehicle Longitudinal Control in Car-Following Scenarios” (Project No. 704/24.04.2026).