Advanced robots are expected to be used in more and more complex and unstructured settings in the future. Robots can now be deployed in factories to repeatedly execute human-designed routines with high robustness. However, to accomplish complex tasks in unstructured settings, the robots must have the capability of reasoning over the task. Task and Motion Planning (TAMP) is a class of methods that combine both high-level task planning and low-level motion planning that enables a robot to reason over both what steps must be taken to finish a task and how to actually do it.

Traditional TAMP literature poses strong assumptions on the class of problems when it is applicable. For example, it is typically assumed that the robot has perfect sensing and execution capabilities, and thus it suffices to find a sequence of motions to finish a task with a real robot. Moreover, many existing TAMP methods focus on single-robot cases and implicitly assume they can scale to multi-robot systems. Such assumptions usually do not hold true when more complex tasks need to be solved (e.g., when execution uncertainty cannot be ignored, or when the task requires coordinating dozens of robots).

This work focuses on relaxing such assumptions and proposes novel formulations and frameworks to address a richer set of problems. We combine the typical TAMP paradigm with statistical models such as Bayesian updates to efficiently reason over the robustness of robotic executions. Beyond execution uncertainties, we also extend our work to consider a more general class of problems, where the robot has various types of knowledge gaps of the world, including object occlusions, unknown objects, etc. To address such challenges, we combine typical TAMP methods with provided execution-level modules, called behaviors, to enable a novel general framework that can discover geometric constraints during planning time instead of real-robot execution time, to finish real-world tasks more efficiently. We also extend a typical TAMP solver to multi-robot problem settings, where we introduce an additional intermediate layer to reason over specific variables, largely increasing the planning efficiency. Lastly, we propose two novel execution frameworks for multi-mobile-robot navigation tasks, combining feedback controller design with sampling-based motion planners and multi-agent path-finding algorithms, to solve such tasks under unknown uncertainties in the high-order dynamic model. Such frameworks are proven to be applicable as execution-level modules to general TAMP pipelines.