Exploration of unknown environments via bayesian optimization for mobile robots

This thesis covers the implementation of an information-theoretic autonomous exploration algorithm for a unicycle robot navigating in unknown environments. The first part of the work is centered around the formulation of the action selection task, which is framed as a Mutual Information maximization problem over a probabilistic log-odds occupancy grid. Because exhaustive evaluation of the sensor field of view is computationally prohibitive, the problem is solved using Bayesian Optimization. By exploiting a Gaussian Process surrogate model, the algorithm efficiently identifies the most informative sensing viewpoints, achieving a significant reduction in computational complexity. The generated optimal target poses are then processed by a hierarchical navigation architecture, which combines global path planning with a local collision avoidance controller to guarantee real-time obstacle-free navigation. Subsequently, to ensure full environmental coverage and escape local minima, an adaptive information threshold and a backtracking mechanism are integrated into the decision loop. The overall framework is initially developed in a custom Python environment and subsequently ported to the ROS2 middleware to perform software-in-the-loop experiments within the Webots simulator. Results are finally demonstrated for several complex mapping scenarios, highlighting how the robot can systematically reduce map entropy and safely navigate around obstacles, while maintaining real-time computational performance superior to the baseline sampling strategies.