UC Riverside

The efficacy and efficiency of mobile robots in real-world applications are challenged by the presence of uncertainty embedded within the environment or the task. This dissertation provides algorithmic solutions for safe and reliable motion and task planning to strengthen the performance of robots while operating under uncertainty. The dissertation's specific focus is threefold: exploration and coverage of unknown regions for information retrieval, handling untrustworthy or inaccurate prior information regarding both the environment and the task during task execution, and navigation in unseen environments.

In the first focus direction, we develop online hex-decomposition-based motion planning algorithms, which address the problem of how to distribute one or a team of mobile robots to explore an unknown, obstacle-cluttered environment cognizant of mobility and sensing capabilities. Here the goal is to fully explore and cover and environment using smooth and continuous paths for robots with non-holonomnic constraints. In the second focus direction, we propose a new stochastic task planning approach that jointly determines optimal task allocation given approximate prior information and expected (yet uncertain) task cost, and decides an optimal stopping time to avoid exceeding given capacities. In the third focus direction, we propose a learning-based method to find feasible, efficient paths for robots to reach the designated task location safely without access to a map of the operating environment. In addition, minimalistic neural network architectures, capable of fulfilling the objective of obstacle prediction, are identified. Together, these three components advance the state-of-the-art in motion and task planning for robot navigation under uncertainty. Moreover, the proposed approaches are scalable and online, hence they can benefit applications that involve real-time motion planning and decision making in the presence of uncertainty.