Design and Implementation of Intelligent Agent Training Systems for Virtual Vehicles

Area of Science:

• Computer Science
• Robotics
• Automotive Engineering

Background:

• Realistic vehicle simulation is crucial for developing and testing advanced driver-assistance systems and autonomous driving technologies.
• The Unity game engine and its Machine Learning-Agents library offer a powerful, flexible platform for creating complex virtual environments and AI agents.

Purpose of the Study:

• To design, simulate, and implement a virtual vehicle with accurate real-world dynamics.
• To develop and train intelligent agents capable of driving the virtual vehicle using imitation and reinforcement learning.

Main Methods:

• Vehicle dynamics simulation in Unity, incorporating realistic elements like motor torque, suspension, and anti-roll bars.
• Development of intelligent agents using Unity's Machine Learning-Agents library.
• Training agents through imitation learning (human expert interaction) and reinforcement learning (reward function optimization).
• Testing agents in simulated highway environments and on procedurally generated, unknown road courses.

Main Results:

• Successful implementation of a virtual vehicle model with high fidelity to real automobile dynamics.
• Intelligent agents demonstrated proficient driving capabilities in both familiar (highway) and novel (spline-based) road conditions.
• Comparative analysis of telemetric data showed performance metrics for both AI-driven and human-controlled virtual vehicles.

Conclusions:

• The Unity platform and Machine Learning-Agents library provide a robust framework for creating sophisticated virtual driving environments.
• Both imitation and reinforcement learning are viable methods for training intelligent agents to control virtual vehicles effectively.
• The developed virtual vehicle and trained agents serve as a valuable tool for future research in autonomous driving and vehicle dynamics.