Let's say someone who's never played golf wants to learn how to swing a club. He could read a book about it and then try it, or he could watch a practiced golfer go through the proper motions, a faster and easier approach to learning the new behavior.
Roboticists face a similar dilemma when they try to build an autonomous machine capable of learning new skills. One approach, as with the golfing example, is to break down an activity into precise steps and then program the information into the robot's brain. This assumes that every aspect of the activity can be dissected, described and coded, which, as it turns out, isn't always easy to do. There are certain aspects of swinging a golf club, for example, that arguably can't be described, like the interplay of wrist and elbow. These subtle details can be communicated far more easily by showing rather than telling.
In recent years, researchers have had some success teaching robots to mimic a human operator. They call this imitation learning or learning from demonstration (LfD), and they pull it off by arming their machines with arrays of wide-angle and zoom cameras. This equipment enables the robot to "see" a human teacher acting out a specific process or activity. Learning algorithms then process this data to produce a mathematical function map that connects visual input into desired actions. Of course, robots in LfD scenarios must be able to ignore certain aspects of its teacher's behavior -- such as scratching an itch -- and deal with correspondence problems, which refers to ways that a robot's anatomy differs from a human's.