Toward Robotic Autonomy in Data-Scarce and Visually Challenging Environments

Abstract

Recent advances in field robotics have allowed robots to be used in challenging unstructured environments (e.g., space, ocean, and natural disaster scenes) instead of sending human explorers since such environments could pose a threat to human lives. However, most field robots are not fully autonomous, relying on communication with human operators. The inefficiencies and constraints of this dependency result in the limited usage of field robots. Hence, enhancing robotic autonomy for unstructured environments is necessary to enable robots to handle large-scale and time-sensitive problems without operators. Among such environments, the underwater domain needs immediate attention since the hydrosphere has high impacts on our lives. Specifically, water covers 71% of Earth’s surface and has the largest ecosystem on Earth. Unfortunately, various human activities have negatively affected the health of underwater ecosystems, which threatens our Earth’s ecosystem as a whole. One such example is the underwater debris problem. Underwater debris has already had damaging effects on the environment (e.g., damaging wildlife habitats), and its long-term effects are predicted to be catastrophic. Existing methods for collecting underwater debris are ineffective, and human exploration and clean-up of the debris are prohibitively risky and cannot adapt to the scale of the problem. This shows it is necessary to deploy underwater robots, often called autonomous underwater vehicles (AUVs), to clean up underwater debris, but the lack of robotic autonomy discourages AUVs from being used for the task. To bridge this gap, this thesis presents our efforts to improve robotic autonomy to make it possible for robots to handle challenging tasks, such as marine debris cleanup, in degraded environments. Part I introduces our efforts to improve robotic underwater object detection, focused on underwater debris, by addressing data scarcity and unique circumstances underwater (e.g., modified shapes of objects over time, degraded vision). In Part II, we present localization and obstacle avoidance algorithms underwater to enable robots to explore a given environment safely and localize underwater objects after detecting them. Even with the improved object detection and navigation methods covered in Part I and II, robots can still struggle with autonomous behaviors due to the extreme challenges that unstructured environments pose in perception and navigation. In cases such as these, human-robot collaborations are still necessary, but robotic capabilities to establish the collaborations are under-investigated and need to be improved. To enable such collaborations, we present algorithms for effective human-robot collaboration in Part III. Lastly, in Part IV, we introduce open-source underwater robotic systems to support robotic research, including the development and deployment process of the research presented in Part I - III. The research introduced in this thesis has been realized and tested on board physical AUVs in various water environments. While our research has mainly focused on the problems in the underwater environment, its robust results in this domain mean that it can be applied to real-world problems in other domains that share similar challenges. Thus, the research has great potential to significantly advance robotic autonomy, bring a broader impact on human well-being, and drive robotics forward for social good.