Autonomous Underground Robotic Inspection
Project lead: Kostas Alexis
Summary: This project aims to investigate the problems of a) sensing, localization and mapping, as well as b) path planning that would enable a small aerial robot to explore the visually-degraded (dark, dusty, smoke-filled) environments of underground tunnels and mines. Simultaneous Localization And Mapping (SLAM) in benign, well-lit, feature-full, and geometrically well-conditioned environments is an established technology using both vision and LiDAR sensing.
However, it remains particularly challenging with only few and weak visual features, self-similar geometry (such as inside a tunnel), or strong presence of smoke, fog or dust. At the same time, despite the significant progress in path planning for exploration, current state-of-the-art presents conservative, slowly exploring behaviors. Furthermore, path planning algorithms rarely account or consider possible inaccuracies and imperfections of the perception system. The result is that when the robot is tasked to navigate in a very challenging environment, researchers resort to more conservative approaches or slower and more careful navigation. An intuitive alternative would be that of inherently accounting for any inaccuracy of the perception system during the planning process and therefore select the next actions of the robot in a manner that is informed about how well it estimates its position and the map of its environment.
For aerial robots to be able to robustly operate and efficiently explore underground tunnel-like and mine environments, and therefore provide rapid means of searching and mapping, the aforementioned technical challenges have to be resolved. To achieve this goal, in this project we envision research in a) specialized sensing and multi-modal sensor fusion, as well as b) geometric reasoning-based exploration path planning. Research investigations will relate to defining the most appropriate sensor suite, developing state estimation algorithms and robust localization and mapping pipelines, as well as studies in sampling-based motion planning and reinforcement learning. In addition, the student teams working in the field of path planning will be asked to design a user interface system that would provide seamless and intuitive utilization of autonomously exploring underground robots by the untrained user.
Student involvement: REU students will begin by observing current University research in subterranean robotic navigation and mapping, and learn how to replicate that work. The students will then select if they will work on the field of perception or planning and will proceed with relevant literature review. A focused discussion that aims to identify the limitations of the current state-of-the-art will follow. Subsequently, the students will design their own solution and approach to the problems of GPS-denied degraded visual environments navigation, exploration and mapping. The implementation phase will follow with students simultaneously building algorithmic, hardware and software development, as well as system integration skills. Finally, students will conduct the necessary experiments with participants on campus and document the results for presentation to the lab with an eye on future relevant investigations.