- Introduces a planner which is called Probabilistic Tree of Roadmaps (PTR) that integrates discrete task planning and continuous motion planning via probabilistic forward search
- It functions over continuous state, and accounts for variability in the motion planning task
- Uses a sample-based tree planner to explore tasks, and a probabilistic roadmap planner to explore the feasible space
- It can actually be difficult in certain situations to determine if a configuration is feasible or not, which can cause problems when planning is done at multiple levels (a high level may think it is feasible, but when examined more closely it may not be)
- Using probabilistic roadmaps (PRMs) are the current state of the art in path planning in large domains, but one issue with the method is it does not have any way of determining if a task is not possible to perform. Once it finds a solution it will return it, but otherwise it simply keeps growing the tree. In practice, a timeout is used to infer if a task isn’t possible, but setting that timeout is generally an ad-hoc method
- “The importance of integrating planning at discrete and continuous levels has long been recognized when studying systems that make and break contact, such as manipulation… This type of planning can be considered as a subset of task-and-motion planning as long as the notion of a ‘subtask’ is construed to subsume the act of achieving a certain contact state.”
- They introduce a STRIPS-like language for doing the planning
- The paper actually doesn’t seem to discuss how to deal with continuous action planning. They mention random sampling and some other things, but don’t say what to use.