Welcome to the last lecture of AI-1. You've almost done it.

Next week, this time, you'll all be free of me. Most of you, probably.

We are looking, we're kind of in a preview mode of the things we'll be looking at next semester.

Namely, planning and acting and searching and so on, in the real world.

When we're no longer so sure that we know everything about the world.

When we're no longer so sure that just because I see something, it means that it is that way.

When we're no longer so sure that every action that we do will actually succeed.

We will have kind of error effects in all the pieces of an agent.

I would like to basically look at what this means from the perspective of planning.

What it means to do planning in partially observable world with uncertain actions, possibly uncertain sensing.

We've already done kind of a thought exercise. What would we have to add to strips?

Or, in this case, in the furniture example, to PDDL.

We kind of have the idea, yes, we need to have perceptions.

Somehow get perceptions into the planning process.

Unless we want to be stuck with what we call sensorless planning.

Planning where we don't need any sensors. Plans that succeed whatever the world is like.

Perceptions is one of the things we need to add.

Actions that actually initiate perceptions. Look at something.

In the end, we also saw that we needed to basically add conditionals to plans.

Where we say, if I perceive this, then do that. If you perceive that, then do this.

Are there any questions so far?

Yes?

We plan for doing sensing.

You can only do conditionals, or usefully do conditionals, if you can sense the environment somehow.

If you can't sense the environment, then what are you basing your conditions on?

We basically have a plan that will include information gathering.

Now, we have this idea of a conditional plan. Let's pick up this ball and run a little bit with it.

What could a conditional plan be?

Remember, we had normal plans, or just sequences of actions.

Pick up this, put down that, unstack, all these kinds of things.

Here, we're extending this idea of plans as action sequences into if C, then plan A, else plan B, phi.

This reminds you of programming. Well done.

We're kind of putting a little bit of program-y stuff into plans.

Makes plans more expressive, and possibly shorter.

If you think about programs, you can think of them as, especially imperative programs,

as action sequences with various little decorations of the form, if then else, or while do, or repeat until,

or give this subplan a name, and then do it 13 times, or something like this.

Programs are actually plans. Very expressive plans.

We're very cautiously basically doing this.

Now, this C here is typically something where we would use sensing in some kind of a way.

A plan would be saying, look at this. If this is reg, then do this. Otherwise, something else.

Now, this is something where you can ask yourselves, why?

We have more plans now. What are the algorithms or the mechanisms that give us a plan like this?

Because we need a new set of algorithms, because our old algorithms are not going to give us these kind of plans.

So how might that work? Well, back to the vacuum cleaner.

If we had a variant of the suck action, you remember, suck up till now, where everything was deterministic,

made a dirty room clean.

And if we had something like a suck action, which is probably a very powerful vacuum cleaner,

which either if it's dirty, you clean the square, sometimes you also remove dirt from the adjacent square.

And unfortunately, if it's clean, sometimes you actually make the carpet that was clean, dirty again.