The amount of literature about “Learning from demonstration” (LfD) is huge. The idea, in short, is that a human operator is moving the robot's arm slowly, the motion gets recorded, and after pressing the replay button the system can do the action – magically – autonomously, even if the environment is a bit different.

What is learning from demonstration exactly?

The working hypothesis is that LfD is the same as "plan recognition". The demonstrated actions are recorded, interpreted and mapped with a given plan notation. That means the AI system is able to identify the plan of the human operator. What speaks against this thesis is that plan recognition is different from machine learning. Because a plan can be recognized without using neural networks or other forms of learning methodology. Another argument against the thesis that LfD is equal to plan recognition is that plan recognition isn't about replaying an action by a robot, but it's more located within the surveillance domain in which actions are monitored but not executed autonomously.

To make things more complicated and blur my original question a bit, I've found lots of different keywords in the literature which have to do with teleoperated annotation, imitation learning and activity grammar. Can anybody help to brighten up the discourse?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.