They tend to collapse into nonsense and hallucinations pretty quickly if you move slightly out of the envelope of the current visual reasoning benchmaxxing.
I'd say virtually all robots you've seen in the real world today rely on classical approaches - you build a rudimentary map, then use classical algorithms to find paths/do area coverage. The robots do no reason or understand what they're looking for, they're more like in-game units. At most there's some bounded, lightweight image classification going on.
LLMs can understand and reason about the world natively. nVidia has a Tesla FSD/Waymo competitor which simply their 7B reasoning LLM but instead of outputting tokens directly, its outputs are fed to a 2B diffusion model that outputs 1.6 second long trajectory for the car, and this is enough for an L2 system. But to make this work, they need the model to run at 10Hz, so they use super high-end hardware to do it (Jetson Thor) and the car is still "blind" for 100ms at a time (they have a parallel classical safety system).
With on-chip LLMs you could run this loop at like 100Hz on a chip that costs a few hundred bucks, rather than 10Hz on a board that costs several thousand.