Learning-based control strategies for safe autonomous robot navigation in unknown environments

This thesis is focused on the implementation of learning-based control strategies for safe navigation of autonomous mobile robots in unknown environments through the use of control barrier functions (CBF). To properly handle unknown environments, the safety function is learned online and in a non-parametric way exploiting Gaussian processes (GP) that can provide a probabilistic representation of the safety function by only using the data coming from the robot’s sensors, with no need of storing past information. The work begins with a theoretical foundation of safe learning for control, with particular attention to the problem formulation and the key concepts of safety controllers and Gaussian processes, which are fundamental to developing the proposed solution. The algorithm is then presented, starting from the Python implementation for initial testing, analysis, and visualization, and subsequently in ROS2 using the ChoiRbot toolbox for more realistic simulations and to allow further development of safe navigation in a distributed multi-robot scenario. Finally, simulation and experimental results are presented and discussed, showing the effectiveness of the proposed solution in enabling safe navigation in unknown environments.