4C: Custom-and-Correct-by-Construction Controller Synthesis using Multi-modal Human Feedback

Abstract

Autonomous vehicle technology has a critical role in enhancing road safety by reducing human errors. The adoption of such technologies is shown to be dependent on user satisfaction, which is influenced by diverse comfort and performance preferences. This paper addresses the challenge of generating custom and naturalistic autonomous vehicle behaviors to meet individual user expectations while ensuring safety. To achieve this goal, we incorporate multi-modal human feedback with pairwise preferences and driving demonstrations into a control synthesis framework. We write safety specifications and express preferences using Weighted Signal Temporal Logic (WSTL), and we employ the PyTeLo toolbox for controller synthesis. Our approach is validated through synthetic experiments and a pilot, small- sized human subject study, demonstrating the effectiveness of integrating multi-modal human feedback for customizing autonomous vehicle behaviors.