This is the interesting stuff.
It's possible the robot ideas will fall prey to the same issues. If you define a complex enough task, maybe the machine will fail to plan what it has to do properly. And trying the tiny tests with a bag of crisps, isn't really all that much of a challenge, from a complexity perspective.
"Take this 40 foot container of ASML Lithography machine parts, assemble and calibrate the machine."
The motion on the Google robot, isn't all that smooth. Boston Dynamics has some software they use for "choreography", that smooths out some of that. Maybe telling the Google robot "how to be smooth", would automatically result in it choreographing what it is doing. Instead of "bump-bump-bumping" the drawer shut. But the bumping behavior likely results from the computer re-evaluating the problem every 0.2 seconds and working out a new set of actions (which includes another tiny "bump"). So perhaps that's an artifact of the method.
Paul