Implementation of Neural Network-based Real-time SLAM Algorithms in a Robotic Simulator for Aerial Robots

The purpose of this thesis is to apply Dense Visual SLAM algorithms to a nano-quadrotor for navigation and mapping. Neural Real-Time SLAM algorithms allow high-quality dense reconstruction in real time. The key idea of this thesis is to implement two algorithms, NICE SLAM and Co- SLAM, both in a real and a virtual environment. In the real-world setting, the Realsense D435i camera is used to capture images to feed the algorithm, providing a real-time 3D reconstruction. In the virtual scenario, a Crazyflie nano-quadrotor is used to map a simulated environment. Here, an RGB- D camera sensor was designed to capture images. The simulative scenario is built on the CrazyChoir ROS 2 Toolbox, where Crazyflies can be easily managed on Webots. Moreover, another important contribution of this work is embedded in the use of simulated environments (e.g., Webots) to test and develop Neural Real-Time SLAM algorithms, allowing for a safe deployment in the real world.