Their biggest problem nowadays compared to humans (and one I hear is in the focus of the AI companies), is that they cannot follow a plan for days, weeks, months, years, while we can. They compensate by being able to be trained much quicker and in many cases better than humans can (up to months instead of years), and once one is trained, any other AI instance of the same or similar model can copy that training almost instantly - that will make a huge difference for robots, when you can mass produce them and install the same "software" in all of them and upgrade it to all. That may also be a problem if those upgrades or the initial software contains glitches, bugs, limits that can be breached, etc. because you have an army of faulty robots potentially going "nuts", as we have seen in one public case.
Intelligence is relative. One could say they are not because they only rehash what they have learned in training and since or what they can find online or in the sources you provide or the prompts, so at least LLMs are language-intelligent and may mimic other types of intelligence based on the training. But they mimic it very well in many cases. Not to mention kids learn by mimicking their parents and people around them too.