Autonomous Navigation and Obstacle Avoidance Stack for the JUNO vehicle

This thesis is the outcome of the activities carried out since July 2023 to the present day in Team H2politO, a team made of students with the aim to design, build and optimize low consumption prototypes in order to participate every year in an international competition, the Shell Eco-marathon. Every technical choice and each component, from mechanics to electronics, derives from a wide collaboration between the team members. The main soft skills that group all of them are the passion and the willpower of defining a target and trying to reach it just with their own forces. A consistent personal growth is therefore experienced putting together different ideas and discussing technical solutions in order to reach the final target: obtain the lowest possible fuel consumption at the Shell Eco-marathon. This thesis wants to represent a document that allows the Team to maintain the know-how built over last months. In fact, every year a partial turnover between team members takes place, so it is fundamental not to lose the experience and knowledge accumulated, transmitting also through to the theses that every year are enhanced. This master’s thesis introduces a novel software stack for autonomous driving, which integrates existing modules for path navigation and obstacle avoidance. The stack was developed by Team H2politO as part of the Shell Eco-marathon’s Autonomous Driving Competition, and designed to be implemented on the team's JUNO vehicle. The objective of the thesis was to develop a solution to complete sectors one and two of the competition: sector one involves autonomous navigation of a vehicle on an empty racetrack, while sector 2 involves avoiding static obstacles while staying within track limits. The novelty of the solution lies in its ability to meet the objective without the need for pre-existing maps or geolocation data. The stack is comprised of two main modules. The first module focuses on lane centering, utilizing a monocular RGB camera and the HybridNets neural network. The network is effective in processing visual data for route planning, complemented by a geometric lateral controller, the Pure Pursuit controller, for maintaining the desired trajectory. The second module focuses on maneuvering around obstacles in the vehicle’s immediate vicinity. The module combines the monocular RGB camera with infrared depth-sensing technology, and is made up of two main algorithms: RTAB-map, a graph-based SLAM algorithm to dynamically interpret the environment, and the TEB Local Planner, a model predictive controller, which facilitates real-time adjustments for navigating around obstacles detected in the environment. Written in Python code, the software architecture is implemented using ROS (Robot Operating System), to communicate within its sub-modules. The integration of the system with the ROS middleware enables the vehicle to process complex data from multiple sensors and implement its navigation and obstacle avoidance strategies efficiently. This thesis outlines the work done in the development and validation of the system in the CARLA simulation environment followed by real-life testing, illustrating the practical application of machine learning, computer vision, robotics, and vehicle dynamics in developing autonomous driving technology.