Usually I put master disconnect switches on my robots just to make working on them safe. I use cheap toggle switches though I'm too cheap for the big red spiny ones.
It's only going to do that if you RL it with episodes that include people shutting it down for safety. The RL I've done with my models are all simulations that don't even simulate the switch.
Which will likely work for only on machine AI, but it seems to me any very complicated actions/interactions with the world may require external interactions with LLMs which know these kind of actions. Or in the future the models will be far larger and more expansive on device containing this kind of knowledge.
For example, what if you need to train the model to keep unauthorized people from shutting it off?