Neuro-Fuzzy Obstacle Avoidance: A Fast Training Dataset Generation Approach Using Low-Cost Sensors

Adaptive Neuro-Fuzzy Inference Systems (ANFIS) are increasingly used for intelligent obstacle avoidance in robotics due to their combination of the learning capabilities of neural networks and the reasoning capabilities of fuzzy logic, providing a powerful approach for handling complex, uncertain environments. This study proposes an ANFIS system for intelligent obstacle avoidance using low-cost ultrasonic sensors. The key contribution is the training database, created by combining model-based data and human reasoning without the need for real sensor data. The database can be tailored to the robot's geometry and sensor features, integrating human-like reasoning by analyzing common and complex avoidance scenarios. It provides fast and suitable input for the ANFIS system, which learns to control the robot's wheel speed to avoid collisions and reach a desired point. The Neuro-Fuzzy approach was compared with two other classic avoidance schemes and assessed in simulation using Cop-peliaSim and Matlab across scenarios of different complexity. The generated dataset allowed training an ANFIS system capable of addressing complex obstacles like walls and maintaining constant velocity and acceleration levels.