Development of space rovers mobility systems for planetary surface exploration and astronaut support with AI methods
This Ph.D. research aims to investigate the use of innovative technologies in planetary ground mobility systems for exploration and astronaut support. The ability of an agent to autonomously explore and navigate an unstructured and previously unseen environment, as is the case for robotic missions to the Moon and Mars, where only low-resolution orbital maps of interest sites are available, is highly desirable. This minimizes the time spent on human-based remote planning while increasing the mission efficiency, allowing the agent to explore larger areas and complete more tasks. In order to achieve these objectives, the robotic agent needs to be able to localize itself in the environment while at the same time generating and storing a map of it for later reuse. Due to the absence of GPS-like localization architectures, the task needs to be addressed by relying only on the robot’s onboard sensors, such as cameras and inertial measurement units.
This approach is known as Visual Inertial Simultaneous Localization And Mapping (VI-SLAM), a popular topic in the literature that is considered solved. However, the multitude of methods presented are prone to fail when dealing with highly unstructured environments, with a scarcity of visual features, and which suffer from extreme visual aliases due to variable lighting conditions.
The main contributions of this research will be twofold. First, an investigation of novel VI-SLAM approaches for the system to be robust and to work reliably in lunar and martian environments will be pursued. For this, traditional computer vision techniques based on environment geometry and appearance, while at the same time exploring novel algorithms based on deep learning to learn long-term stable features for data association. Furthermore, methods for actively planning and control of robot motion to decrease the uncertainty in the generated map will be studied. These will be addressed with the modeling of a decision-making process based on reinforcement learning for balancing exploration and information exploitation. The aim is to increase the chance of performing place recognition and loop closure in difficult environments and to smartly generate and maintain a map for long-term missions, while at the same time satisfying the strong power and computational constraints typical of space agents.
Department: PhD in Eletrical, Eletronics and Communication Engineering
Supervisor: Prof. Marcello Chiaberge