This is a Section of the book "Robots learning from human teachers" consists of elements that robot learners should take into consideration. These elements are inspired by human and the way humans interact. It basically, integrates social learning into Learning from Demonstration.
2. Inspired from Human Learning
● Insight to advance learning systems
● Interaction between human and robot
● Design elements for robots using Learning
from Demonstration
○ Social Interaction
○ Motivation for Learning
○ Transparency
○ Question asking
○ Scaffolding
○ Directing attention
○ Online Vs Batch Learning
4. Transparency
● Teacher needs to know what robot knows
● Robot should display what it already knows
● Ways to communicate about this knowledge
from robot to human.
Question Asking
● Posing a question
● Getting information from
answer
● Asking relevant questions
5. Scaffolding
● Breaking down tasks in simpler tasks
● This increases efficiency for re-use
● Parametrizing previously learned skills to
generalize
Directing Attention
● Focus learning objective
● Used for feature selection in
machine learning
● Simplify learning in complex
environments
6. Online Vs Batch Learning
● Producing all training data as once and then training it on the model.
● Batch learning starts at the end of each learning session or once new demonstrations are
acquired.
● In online learning or active learning the process of acquiring the data and training the
model are done simultaneously.
● Choice between two is required to decide the flow of interaction and the way of integrating
new training data in the model.
7. Review Paragraph
This Section consists of elements that robot learners should take into consideration. These elements
are inspired by human and the way humans interact. It basically, integrates social learning into
Learning from Demonstration. These elements might or might not be useful for robots learning from
demonstration. Hence, this section gives a brief overview of elements that can be considered. Social
learning and learning from motivation are common types of learning inspired by humans, which can
be very useful in Learning from Demonstration as they can be easily converted into code blocks which
gives similar results. Having transparency in robot learning is necessary to make the process of
learning clearer and avoid singularity. Similarly, scaffolding helps in breaking down complex task to
simple tasks. Question asking is another way in which complex tasks can be done by asking simple
and relevant questions. Humans use a very common method of learning in which they try to direct
the attention of the learner by sound, motion, colour, light etc. This can also be a good way for robot
learning as it can have the capabilities to sense these factors. The last one mentioned in this section is
the online vs batch learning which is again an effective way of robot learning.
Editor's Notes
Interaction between human and robot
This may or may not be necessary depending on the application. For some applications recognizing and reacting to human gestures can make a robot more human like and affectionate. Hence interaction is necessary for such application of learning from demonstration. For some tasks no human interaction is needed. Hence, design of learning systems depend on the nature of the tasks.
Social interaction between robot and human could be aided by identifying and providing social cues, these social cues include facial expression, vocal tone, body language, body posture, gestures and proximity. Like in the telenursing robot the robot can also identify a person as a patient by his body language and his facial expression if the person is in pain. Hence identifying the social cues and surely improve learning from demonstration.
Motivation for learning
Robots can be given motivation by reward function as it is given to children using reinforcement learning or it can be guided by human. But if in initial stage it is guided and then let it learn by using the reward function to improve the task that could achieve a better learning curve.
Transparency:
In order for the human to teach the robot a task, it needs to know what prerequisites the robot already has and what can the human teacher start with. Hence transparency in the learning process is necessary. In order to achieve this transparency, first teacher needs to understand what knowledge the robot has. This can be achieved only if there is a display system of the robot knowledge to human teacher. Example of transparency can be such that, lets say we want the robot to pour water from a bottle to a cup. Then if we have transparency we can figure out if the robot already knows how to grasp objects before starting to teach it.
Question asking:
In order to implement new skills on the robot it is needed to program them on it every time. But this is very tedious, hence the question asking method is useful here. The robot Simon Georgia tech’s is programmed to good questions in order to learn a skill. Designing proper questions is necessary to learn the correct skill with ease.