Working in a few related fields, they are already being combined to some extent. We have machine perception, where the bot can often fine objects in the world around them, and so things like pick them up and move them around. on the other end you have all these deep learning methods that can help simplify large data sets, and that helps make it easier to find things more reliably. The problem is they are all probabilistic... The machine will easily confuse objects (a dog for a loaf of bread), and then it can misjudge the world around it an unintentionally break things or hurt people.

There also practical issues, power, sensors, are all still in the early days, and largely inefficient and otherwise expensive. Most of the bots only have minutes of runtime before they need to charge again.

intelligent and helpful are tall orders given all that... combining the above is still wildly off from intelligently working side by side with a human.

I think a whole new operating system needs to be invented that sits above all this, and orchestrates things... Receiving commands without confusing intent, interacting with the world without serious mistakes, and working with objects it can reliably identified.