Adaptive Navigation System for an Autonomous Vehicle in a Goal-Oriented Environment

In the context of autonomous navigation, the development of systems that enable vehicles to operate independently in controlled environments is a crucial step toward advancing autonomous technology. This work presents the design, implementation, and validation of a navigation system for autonomous vehicles using NeuroEvolution of Augmenting Topologies (NEAT). The primary objective was to create a vehicle capable of navigating a 2D map with a defined starting point and target. Virtual sensors enable the vehicle to identify navigable paths and boundaries. Distance metrics such as Euclidean, Manhattan, and Chebyshev were employed as reward systems, continuously calculating agent positions. The closer the vehicle is to the target, the higher its fitness score, forming the basis of the fitness function. A forced reinforcement acceleration method was designed and implemented to ensure progress when the vehicle's speed fell below 0.1, preventing it from becoming stalled. Validation tests were conducted to evaluate the system's performance under varying conditions. Results demonstrate that the autonomous vehicle can navigate the map effectively, improving its fitness score in each generation depending on the distance metric used. Chebyshev performed best in obstacle-free environments, while Euclidean excelled in the presence of obstacles. The forced reinforcement method significantly reduced the time required to achieve the target fitness. These findings provide valuable insights for researchers aiming to develop NEAT-based navigation systems for autonomous vehicles.