0:00

So, in the last lecture, we saw that, even for rather un-, scary looking

environments. Let's say that we had, slight non

convexities, or even extended convex obstacles.

the dynamic duo of behaviors, go to goal and avoid obstacle,

no longer was enough to negotioate these environments.

And we saw that we really needed some way of flowing a long the boundary of an

obstacle in order to be able to get around in the world.

And today, I want to talk about this missing piece.

The missinig behavior which is a wall following behavior.

And what I want to do now is really just introduce it and then in the next lecture

introduce it in a more systematic way as to induce mode when you're trying to

avoid sliding or avoid type 1 Zeno. So, here is the question.

How do you actually follow walls? Or when I say walls what I really mean are

boundaries of. boundaries around obstacles.

So let's, let's start simple. Here is, as always, an obstacle and I

have a disk around it where, this is my safety disk.

Well, avoid obstacle is going to tell me, you know what, I want to move straight

away from this From this obstacle. Well, how do I , which direction do i go

in, in order to maintain a constant distance from this obstacle? Well what I

need to do is, I need to flow, right, along the boundary of this, this region.

Well, what this means is that here, I need to go in this direction.

And what I really need to do then is take The direction, given to me by my avoid

obstacle behavior, and flip it down, like this.

And this is a negative pi over 2, or a negative 90 degree flip.

So, that's what I need to do. I need to flip it down by 90 degrees,

like this. And now, I'm going to call this FW, for

Follow-wall, and follow-wall is now, simply, like I said, a flip down of 90

degrees, which this curious looking matrix here is actually achieving.

And I may want to scale it, so I may have some scaling that's making this Longer or

shorter, depending on, I don't know. Whatever concerns you may have.

Now, I should point out that, this matrix here, I can write as a rotation matrix,

that corresponds to a pi over 2. Sorry, negative pi over 2 flip.

so I can actually write this as alpha, which is the scale, times the rotation

matrix, that is, the flip, times u, avoid obstacle, because that is ultimately the

behavior that I'm, that I'm flipping. So, this tells me that the way to move

around this obstacle, is to simply take my avoid obstacle behavior, and flip it.

That's as simple as that. So, there is a problem though.

I can actually flip in 2 different directions.

Not only can I go like this, I can also go like that.

So 1 question 1 should ask of course is which direction should I go in, but the

other question is how do I even mathematically describe these 2 options?

Well, what I have, what we saw before was, we called it U clockwise follow

wall, and the clockwise follow wall corresponds to a negative pi over 2 flip

to avoid the obstacle. A counter clockwise follow wall, which

means follow it in the other direction, well that is simply a pi over 2 flip of

the void obstacle behavior. So if I introduce this thing that I've

already called the rotation matrix, this is simply an operation that does this

flip over vector. If I have a vector here.

V, and I want to flip it by phi. This new thing here, is simply r of, it's

not phi, it's theta. R of theta times v.

This is how I get this new, this new rotated vector.

So the rotation matrix is cosine theta negative sine cosine.

That's the form of this linear operator that's flipping vectors at certain angle.

Well what we've already seen is that the clockwise follow-wall it's a negative pi

over 2 flip so this behavior is simply equal to alpha times, this matrix times,

uAO. Now the counterclockwise, while I can

write that in the same way, it's just a pi over 2 flip instead of a negative pi

over 2 flip, which gives me the following expression.

So, I actually have very, very clean expressions for how I should make my

flip, meaning how I should proceed along the

boundary of the obstacle. And this works even if their obstacle is

not just a point, because all I really did was I took a void obstacle, a void

obstacle behavior and, and flipped it. Now we don't know which direction to go

with them. So lets say that we have this situation,

right. [SOUND] I'm going towards the goal,

and then let's say when I'm here, I now want to switch to a follow-wall behavior.

Which direction to choose? it's not entirely clear.

Right? I, if the robot only sees what's going on around here, for instance, how

does the robot know what the world looks like? We haven't built a map of anything.

So, you could either go clockwise, which in this case turned out to be the

long way around, or you can go counter-clockwise, but it's really not

clear how we should make this choice. But here's one thing one can observe, is

that go to goal here is wants to move you kind of upwards a little bit.

So maybe we can let go to goal make the choice for us.

Go to goal once the drive was a little bit in this direction, not at all in that

direction. So that's the idea we have here at, that

there is no obvious answer, there is not then obvious right choice given what

local information the robot has. The robot doesn't know globally what the

world looks like. So given that we only know locally what's

going on may be we can lack this goal to goal behavior.

Decide which direction we should go in. So let's, let's try to make this a little

bit more mathematically precise, what it means to let the goal to goal behavior

decide. Here, goal to goal, the angle was wrong

here. Goal to goal wants to go in this

direction. Right? So here, goal to goal wants to

drive us up. Limit up, so may be we should take

follow-wall up. So this would be u clockwise.

Well, u counter clockwise or u follow-wall counter clockwise wants to go

down. And note, that this angle is less than pi

over 2, but this angle, the absolute value of it, is greater than pi over 2.

So we should be able to look at this angle to determine.

Where we should go, but lucky for us there's a very simple way of checking

these angles. If I have two rectors v and w I can take

what is called the inner product of those t which is simple v transposed to w.

Well that's equal to the absolute values of, or the, the length of vectors times

cosine of this angle in between them. And if this angle is less than pi over 2,

then cosine is going to be positive. If it's greater than pi over 2, then

cosine is going to be negative. So all we need to do.

It's to look at the sine of this thing, to figure out whether or not we're kind

of pointing in the right direction. So, here is a test.

Check go to goal inner product with follow wall clockwise.

If that's positive, we should go clockwise.

So, in this scenario, again. Again, we're pointing in this direction,

here is the follow wall, so this, go to go.

Here is u clockwise. This angle, is less than pi over 2, which

means that, this situation is satisfied. Here is u counter-clockwise, so this

angle is greater than pi over 2. So, that means that the other condition

is not satisfied. So what we actually have is a local

check. We compute the direction that two

different follow-wall behaviors would take us in.

Then we just take this inner product with go-to-goal behavior, and we check the

sign. And whoever comes up with a positive sign

is the winner. and that's the direction we choose to go

in. Now, there is a problem with this of

course. And that's that if you remember our

labyrinth, or our little maze here, right.

If I'm following wall, you know what, I'm going to follow it all the way around.

And I don't want to keep going indefinitely around and around, right.

So then I'm just spiraling. At some point, if here's the goal.

I want to release and go towards the goal.

So one thing that we don't know yet is. When do we actually release the

follow-wall behavior? When do we stop following walls and start approaching

goals again? The other question is, you know what, we had avoid obstacle we had

the rotation matrix, that flips either up or down, depending when we're going

clockwise or counterclockwise. And then I have this little alpha there

Is there some natural or systematic way on which we can think about the alpha?

And that is going to one One of the things we will discuss in the next, the

next lecture.