When a system (say, transformer model) is trained to output full plans conditional to a situation — would one call this planning?

Task and Motion Planning (TAMP) became a standard problem formulation within robotics to describe the complex behavior we strive for. However, despite the word Planning in TAMP, it doesn’t mean that one has to use planning methods in the traditional sense — it rather characterizes the behavior we want to see. In this talk I will discuss own work on TAMP, covering both, planning (i.e., optimization & sampling) and learning methods, and try to clarify these characteristics we want to see. I will also spend some time discussing whether “getting things to work” is our only goal, and the “call for more data” the solution.