Unless the only thing you want to do with the robot is talk, you need to do a lot more reasoning and execution planning first (= multiple LLM round trips; tool

Unless the only thing you want to do with the robot is talk, you need to do a lot more reasoning and execution planning first (= multiple LLM round trips; tool calling) before you even know whether talking is the correct action to take. So the naive time-to-first-chunk estimate will be way off.