So if you go through perception and understanding,

if you take understanding as generating a homotopy

between bird and mind, so we would have done the first two

steps in it.

And then let's say something about action.

And I do not want to repeat the part that you are responsible

for setting a target function.

And so I want to speak about the mechanics of the action.

And again, this is a type of autoencoder procedure.

Action means you have an idea in your mind,

then you would like to act in the world.

Then you receive an answer back to your mind.

And hopefully, the reception is in the direction

of what you want to have there.

This is not clear.

So think about this example with driving a bicycle there.

You can say, I want to drive the bicycle,

but you are not able to control this machinery.

And therefore, you fall down to the ground.

Or let's take the example with the bow and arrow spot

there on the right side of the picture here.

You want to shoot on the center part of this target there.

But if you can do it, that's a completely different question.

It depends on how good you are able to emit

the control to the world.

And then afterwards, you receive the answer,

yes, you were able to do so or not.

And from the beginning on, there is no guarantee

that you can do it in a correct way.

And when I was younger, I did do archery.

I can guarantee to you that you will not

do it from the beginning on.

And so the point is, again, we have an emissive part.

And again, we have a receptive part.

Now, this is starting from the mind going through the world

and going to the mind again, which

means you can reuse the emissive and receptive part

that you have trained before for the perception

to do so.

And nevertheless, again, here, you have to do the training.

If you want to do something very good, you have to train it.

And the good news about this is it works even if you do not

have from the beginning on a perfect model

of the environment.

And so the thing is by the learning, you can do it.

And again, you have your auto-encoder there.

And so the topic here is perception and action

on both sides fluctuating movements, information flows

forward and backward.

And you have to train this as fluctuations.