IMO, the struggles with AI in robotics exposes where the current AI bottlenecks are more clearly.
Specifically, we don't have a training data corpus for any of our AI robots, enough to help them perform generalized task behaviors. So we have to go out and create them / use non training techniques, which crucially cannot leverage the existing learning / movements of humans accomplishing things.
ChatAI's training data of "generalized text on the internet" mirrors exactly what it can do right now — know a bit about everything. But there lacks computer-readable corpuses for the things that are truly useful — 1) efforts that lead to a task accomplishment and 2) better-than human efforts that lead to better-than-human output.
The key question is this: can AI reach beyond the human dataset that is currently powering it, crucially, at low(ish) cost? Or will AI always be fed training data? If we think the former, then we have to admit our algos aren't quite there yet and this iteration of AI is selling that dream.
But even with our current data-fed AI algos, we have not yet put a lot of effort into the root of the challenge - how to quickly build large sets of high quality training data. I think there can be optimization here (over a period of 10-20 years) on a similar scale of going from million dollar computer mainframes to hundred dollar laptops. If high quality training data for specialized efforts were indeed available at a low cost, then the AI revolution will effectively deliver the dream it promises - just wait for Moore's law.
The first step to solving that is an extremely polished LLM. You can see the proof of concept with SORA and other image-to-text efforts on how quality training data is created with the help of LLMs.
But of course, "LLM generated great training data" doesn't draw headlines, so it may just look like nothing is happening for a long, long time, until you see AI task bots pop up everywhere.