Deep Reinforcement Learning based mapless navigation and control of mobile robots.

Abstract

This thesis presents a pipeline for mapless navigation of mobile robots, where decision- making and control are handled in separate stages. A Deep Reinforcement Learning (DRL) agent, trained with artificial neural networks, generates velocity commands that allow the robot to reach a goal while avoiding obstacles, using only onboard sensor data. These commands are then passed to a fuzzy Takagi-Sugeno (T-S) controller, which ensures accurate and robust trajectory tracking. In the single-agent case, the DRL-based navigation is compared with a classical navigation approach. The framework is further extended to a multi-robot setup, demonstrating decentralized coordination in shared environments. Simulation results validate the effectiveness and adaptability of the proposed pipeline.