Multi-Task End-to-End Self-Driving Architecture for CAV Platoons.

Area of Science:

• Robotics and Artificial Intelligence
• Transportation Engineering
• Computer Science

Background:

• Connected and autonomous vehicles (CAVs) offer potential benefits like reduced emissions, improved safety, and enhanced comfort.
• CAV platooning, with close inter-vehicle distances, promises further gains in energy efficiency, space utilization, and reduced travel times.
• Limited research exists on self-driving algorithms specifically for urban CAV platoons.

Purpose of the Study:

• To propose a novel end-to-end deep learning architecture for self-driving CAV platoons in urban environments.
• To develop a multi-task model capable of controlling both leading and following vehicles within a platoon.
• To integrate sensor data with vehicle-to-vehicle (V2V) communication for robust control.

Main Methods:

• An end-to-end deep learning approach using a single neural network to predict steering and throttle/brake commands.
• Input data includes front-facing camera images augmented with V2V communication information.
• Training and testing conducted in a simulated urban environment with dynamic traffic, comparing against recurrent neural networks and transfer learning.

Main Results:

• The proposed architecture demonstrated successful control of CAV platoons in simulation.
• A two-vehicle platoon achieved 67% completion rate in the training environment and 40% in an unseen environment.
• V2V communication effectively eliminated causal confusion for following vehicles, a common issue in end-to-end systems.

Conclusions:

• The developed end-to-end architecture is a viable solution for urban CAV platooning.
• The integration of V2V communication significantly enhances the reliability and safety of autonomous platooning.
• This research paves the way for more efficient and safer urban transportation systems.