Robotics Needs Non-Classical Planning
Classical planning has developed a powerful set of abstractions and assumptions that enables the study of the underlying characteristics of real world problems. While these abstractions and assumptions are beneficial in academic research, they prove to be a barrier against the direct application of classical planning to real world problems and systems. Similarly, non-classical planning approaches have been developed, constructing the necessary bridges between classical planning's assumptions and the hard truths of operating in the real world. These techniques remove many of the assumptions that simply do not hold while operating in the real world. They remove assumptions such as: a fully known initial world state, fully known future world states and unbounded, uninterrupted planning time.
This dissertation makes two contributions. First we show how uncertainty in the world model can be addressed through a non-classical planning algorithm called hindsight optimization. We consider two realistic sources of uncertainty: temporal uncertainty and open worlds. The second contribution is applying abstractions and techniques from the heuristic search community to motion planning. We demonstrate the power of abstraction in a complicated task and motion planning problem with temporal constraints. We then show how combining high level discrete reasoning, characteristic of heuristic search, can aid lower level sampling-based motion planning resulting in faster solving times and better solution quality.