Two-Level Architecture: Planning-Control

A robot needs solutions at two timescales: long-term planning (0.1-1 Hz: “clean the kitchen” → sequence of subgoals — semantics, dependencies, multi-step plan) and VLA-level control (1-10 Hz: specific operation — smooth control signals). The separation simplifies complexity: planning answers “what to do”, control answers “how to do it”. This is structurally similar to fast/slow thinking in speech, but the cut is different: “plan holistically” versus “execute in real-time”. The real-time requirement doesn’t disappear, but is shifted to the control level and mitigated by breaking actions into blocks — trading reactivity for smoothness.

Related: [VLA models], [Thinking, Fast and Slow architecture]