Li, Tianyi (University of Minnesota), Wang, Shian (The University of Texas at El Paso), Shang, Mingfeng (Univerisity of Minnesota), Choi, Seongjin (Korea Advanced Institute of Science and Technology, KAIST), Stern, Raphael (University of Minnesota)

A Customizable Neural Network Based Framework for Autonomous Vehicle Control with Human-Guided Learning

Scheduled for presentation during the Invited Session "Large-scale Smart Mobility" (FrAT2), Friday, September 27, 2024, 11:50−12:10, Salon 5

2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), September 24- 27, 2024, Edmonton, Canada

This information is tentative and subject to change. Compiled on October 3, 2024

Abstract

This study introduces a novel control framework for adaptive cruise control (ACC) in autonomous vehicles (AVs), utilizing Long Short-Term Memory (LSTM) networks and physics-informed constraints. The LSTM component captures complex vehicle dynamics and temporal dependencies, while physics constraints ensure realistic operational limits. This framework supports customization of control objectives, allowing for the integration of various performance metrics to achieve specific goals such as reducing speed variation and enhancing traffic flow.

A distinctive feature of this framework is the implementation of human-guided learning, where a human-controlled lead vehicle profile informs the training process. This human-in-the-loop approach allows the controller to adapt to real-world driving patterns and complexities effectively. Extensive simulations demonstrate the framework's adaptability across different scenarios, highlighting its capability to maintain safe vehicle spacing, ensure smooth speed profiles, and optimize energy efficiency. The results highlight the controller's ability to maintain safe spacing while ensuring smooth speed profiles and optimizing energy efficiency, depending on the chosen customization settings. This research paves the way for the development of more intelligent, safe, and efficient ACC systems that can be tailored to various driving conditions and user preferences while benefiting from the knowledge and experience of human drivers during the learning process.