My Machine Learning professor at University of Texas was hugely into Reinforcement Learning. He and one of his graduate student had developed this kind of system many years ago. Basically, only using a simple video camera, the computer could identify correctly all sub-activities involved in making a peanut butter sandwich; such as 'open jar', 'jelly on bread', 'jar closed', etc. We were shown the video during first class of the course and it was pretty cool.
The MSR guys are doing interesting stuff but this is nothing groundbreaking.