Building a robot is hard. Teaching a robot is even harder. But teaching a robot to do your teaching for you might just be within the realm of possibility.

If you're building a robot to perform a specific task—like opening a door, for example—there are only a few ways to teach it. Most robots learn via motion planning, where a programmer specifies each movement of the robot's motors. Some robots can also learn by watching and mimicking a human performing the same task.

Both of these methods have their drawbacks. Mimicry is a fast method that serves a robot well when performing one specific task but costs the robot adaptability. If a small detail is changed—such as the type or position of a doorknob—the robot has to relearn the entire exercise from scratch.

Motion planning has the opposite problem: many different conditions and situations can be programmed into the robot, but this takes a considerable amount of time and effort. If you want the most versatile door-opening robot possible, you need to spend countless hours programming ways to open dozens of types of doors.

A team of researchers at MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL) have developed a technology that combines these two teaching methods to take advantage of their strengths while avoiding their weaknesses.