Attention-based Deep Reinforcement Learning for Autonomous Robot Navigation

The pervasive integration of autonomous robots across diverse sectors, from healthcare to manufacturing, underscores the critical need for advanced algorithms that enhance autonomy and precision. Traditional robotic systems often rely on task-specific algorithms or external sensing infrastructure to achieve high performance, which limits their flexibility and adaptability. This thesis addresses these challenges by developing and adapting an attention- based Deep Reinforcement Learning (DRL) algorithm for the TIAGo robot, aimed at enabling autonomous navigation using only onboard sensors. Building on prior work originally designed for the Turtlebot 2, the proposed approach introduces several modifications to account for TIAGo’s structural and perceptual constraints. These include the integration of RGB-D data, customized preprocessing pipelines, and a redesigned reward function. The system was evaluated in simulated environments containing both static and dynamic obstacles. Results demonstrate that the proposed method achieves robust navigation and outperforms baseline approaches in scenarios where LiDAR alone is insufficient, such as indoor environments containing tables. Furthermore, a compara- tive analysis of DRL algorithms shown that Soft Actor–Critic (SAC) provides smoother and more reliable navigation than Deep Deterministic Policy Gradient (DDPG), owing to its inherent stability and exploration capabilities. This work highlights the potential of attention-based DRL models for real-world mobile robot navigation and provides a foundation for future research on safe, efficient, and socially aware robotic systems.