Title:	Contextual Multi-Objective Ensemble-Based Control in AI-Based Autonomy
DNr:	Berzelius-2026-227
Project Type:	LiU Berzelius
Principal Investigator:	Hazem Torfah <hazemto@chalmers.se>
Affiliation:	Chalmers tekniska högskola
Duration:	2026-08-18 – 2027-03-01
Classification:	20208
Homepage:	https://starlab.systems
Keywords:

Abstract

This project focuses on developing and evaluating advanced runtime monitoring and decision-making frameworks to ensure safety and adaptability in AI-based autonomous systems. Machine learning-based controllers are increasingly deployed to solve complex decision-making tasks; however, their performance can degrade in unfamiliar environments or dynamic contexts, such as adverse weather and unpredictable traffic, raising critical safety concerns. While traditional ensemble methods enhance robustness by aggregating predictions across multiple controllers, standard voting or averaging mechanisms often dilute the specialized strengths that individual controllers exhibit in different operating contexts. To address this limitation, we investigate new monitoring frameworks for ensemble AI-based control that guarantee safety while selectively leveraging each controller's context-specific capabilities. In addition to safety, the framework explicitly addresses competing objectives, such as driving efficiency and passenger comfort, by optimizing trade-offs across these performance metrics. Neural network controllers will be evaluated within the CARLA driving simulator to test vehicle-level decision-making and runtime safety monitoring across complex urban driving scenarios.