I am working in robotic domain, Robot Learning from Demonstration (RLfD). However, it is worth mentioning that there is analogy between social learning in developmental psychology and Robot Learning from Demonstration. I am mostly interested in very low level RLfD, i.e. in trajectory level. Therefore given a set of trajectories of a specific task I am interested to understand difference between imitation learning and emulation.
Thank you for your responses. I would like to know your opinion on the definition of imitation learning and emulation in the following example.
Assume that a person (the trainee) wants to put a spoon of sugar from cup A in cup B, where A and B are on a table. If the trainee was to learn by mimicking, it would follow the exact same path as the model did, and would therefore be successful so long as neither cup A nor cup B moved. On the other hand, if the trainee was to use imitation learning, they would be able to successfully accomplish the task no matter where the cups were. In this case, they are able to extract a task model given a set of task demonstrations, recognize the state of the goal, which are positions of cups A and B, and map the observed behavior to the new states of the goal to accomplish the job. Gaussian Mixture Model/ Gaussian Mixture regression, motor primitives for imitation learning and estimating a nominal path have been developed to allow a robot to replicate user movement behavior. In a scene cluttered with objects the nominal model may no longer be feasible, so the trainee uses emulation learning to infer from features of the scene how and when to deviate from the nominal model.
In this case, emulation would extract 1) the conditional relationship. i.e. the scoop out sugar from cup A first and then put it in cup B, 2) the result of scooping from cup A, 3) the result or the fact that the same spoon is put in cup B. The learner will try to choose from possible behavioral solutions to achieve the goals. Apparently children do this before learning to imitate. If it is done using an imitation paradigm, the learner will focus on segments of demonstrations and try to build behavioral models to accomplish 1) reaching to cup A as a statistical summary of demonstrations, 2) scooping from cup A, 3) balancing during transportation to cup B, 4) dropping sugar in cup B. So, imitation encapsulates emulation in some sense. Perhaps that is why children use emulation paradigm first.
In any case, we really don't know what people watch in imitation. It can be the state transitions in end point or join space, or it can be in a more hidden space like in a reward landscape. That leads to things like attention control. Still open for studies. One of our papers titled "Full-Body Postural Control of a Humanoid Robot with both Imitation Learning and Skill Innovation" just got accepted in the Journal of Humanoid Robotics. You might want to have a look once it is out.
- Badges
- Science topic
- Similar topics
- Psychology
- Developmental Psychology