John-Green-bot: Hi, I’m John-Green Bot and welcome to Crash Course AI!

Today we’re learning about meeee!!!!

Jabril: Hey!

This is my show!

John-Green-bot: Uh oh… Jabril: It’s ok John Green Bot, we can do this intro together.

Robotics is a broad topic, because it’s the science of building a computer that moves and interacts with the world (or even beyond the world, in space).

John Green Bot: So today we’re going to talk about robots, like me, and what makes us tick!

INTRO Some of the most exciting AIs are robots that move through the world with us, gathering data and taking actions!

Robots can have wings to fly, fins to swim, wheels to drive, or legs to walk.

And they can explore environments that humans can’t even survive in.

But, unlike humans, who can do many different things.

Robots are built to perform specific tasks, with different requirements for hardware and for learning.

Curiosity is a pretty amazing robot who spent 7 years exploring Mars for us, but it wouldn’t be able to build cars like industrial robots or clean your apartment like a Roomba.

Robotics is such a huge topic that it’s also part of Computer Science, Engineering, and other fields.

In fact, this is the third Crash Course video we’ve made about robots!

In the field of AI, robotics is full of huge challenges.

In some cases, what’s easy for computers (like doing millions of computations per second) is hard for humans.

But with robotics, what’s easy for humans, like making sense of a bunch of diverse data and complex environments, is really hard for computers.

Like, for example, in the reinforcement learning episode, we talked about walking, and how hard it would be to precisely describe all the joints and small movements involved in a single step.

But if we’re going to build robots to explore the stars or, get me a snack, we have to figure out all those details, from how to build an arm to how to use it to grab things.

So we’re going to focus on three core problems in robotics: Localization, Planning and Manipulation.

The most basic feature of a robot is that it interacts with the world.

To do that, it needs to know where it is (which is localization) and how to get somewhere else (which is planning).

So localization and planning go hand-in-hand.

We humans do localization and planning all the time.

Let’s say you go to a new mall and you want to find some shoes.

What do you do?

You start to build a map of the mall in your head by looking around at all the walls, escalators, shops, and doors.

As you move around, you can update your mental map and keep track of how you got there -- that’s localization.

And once you have a mental map and know the way to the shoe store, you can get there more quickly next time.

For example, you can plan that the escalator is faster than the elevator.

The most common way we input that data is with our eyes through perception.

Our two slightly different views of the world allows us to see how far away objects are in space.

This is called stereoscopic vision.

And this mental map is the key to what many robots do too, if they need to move around the world.

As they explore, they need to simultaneously track their position and update their mental map of what they see.

This process is called Simultaneous Localization And Mapping, which goes by the cool nickname SLAM.

But instead of eyes, robots use all kinds of different cameras.

Many robots use RGB cameras for perception, which gather color images of the world.

Some robots, like John-Green-bot, use two cameras to achieve stereoscopic vision like us!

But robots can /also/ have sensors to help them see the world in ways that humans can’t.

One example is infrared depth cameras.

These cameras measure distances by shooting out infrared light (which is invisible to our eyes) and then seeing how long it takes to bounce back.

Infrared depth cameras are how some video game motion sensors work, like how the Microsoft Kinect could figure out where a player is and how they’re gesturing.

This is also a part of how many self-driving cars work, using a technology called LiDAR, which emits over 100,000 laser pulses a second and measures when they bounce back.

This generates a map of the world that marks out flat surfaces and the rough placement of 3D objects, like a streetlamp, a mailbox, or a tree on the side of the road.

Once robots know how close or far away things are, they can build maps of what they think the world looks like, and navigate around objects more safely.

With each observation and by keeping track of its own path, a robot can update its mental map.

But just keep in mind, most environments change, and no sensor is perfect.

So a lot goes into localization, but after a robot learns about the world, it can plan paths to navigate through it.

Planning is when an AI strings together a sequence of events to achieve some goal, and this is where robotics can tie into Symbolic AI from last episode.

For example, let’s say John Green Bot had been trained to learn a map of this office, and I wanted him to grab me a snack from the kitchen.

He has localization covered, and now it’s time to plan.

To plan, we need to define actions, or things that John-Green-bot can do.

Actions require preconditions, or how objects currently exist in the world.

And actions have effects on those objects in the world to change how they exist.

So if John Green Bot’s mental map has a door between his current location and the kitchen, he might want to use an “open door” action to go through it.

This action requires a precondition of the door being closed, and the effect is that the door will be open so that John Green Bot can go through it.

John Green Bot’s AI would need to consider different possible sequences of actions (including their preconditions and effects) to reason through all the routes to the kitchen in this building and choose one to take.

Searching through all these possibilities can be really challenging, and there are lots of different approaches we can use to help AIs plan, but that would deserve a video on its own.

Anyway, during planning we run into the third core problem of robotics: manipulation.

What can John Green Bot’s mechanical parts actually do?

Can he actually reach out his arms and interact with objects in the world?

Many humans can become great at manipulating things (and I’m talking about objects, not that force powers stuff).

Like, for example, I can do this but it took me a while to get good at it.

Just look at babies, they’re really clumsy by comparison.

Two traits that help us with manipulation, and can help our robots, are proprioception and closed loop control.

Proprioception is how we know where our body is and how it’s moving, even if we can’t see our limbs.

Let’s try an experiment: I’m going to close my eyes, stretch my arms out wide, and point with both hands.

Now, I’m going to try to touch my index fingers without looking.

Almost perfect!

And I wasn’t way off because of proprioception.

Our nervous system and muscles help our body’s sense of proprioception.

But most robots have motors and need sensors to figure out if their machine parts are moving and how quickly.

The second piece of the puzzle is closed loop control or control with feedback.

The “loop” we’re talking about involves the sensors that perceive what’s going on, and whatever mechanical pieces control what’s going on.

If I tried that experiment again with my eyes open instead of closed, it would go even better.

As my fingers get closer to each other, I can see their positions and make tiny adjustments.

I use my eyes to perceive, and I control my arms and fingers with my muscles, and there’s a closed loop between them -- they’re all part of my body and connected to my brain.

It’d be a totally different problem if there was an open loop or control without feedback, like, for example, if I closed my eyes and tried to touch my finger to someone else’s.

My brain can’t perceive with their eyes or control their muscles, so I don’t get any feedback and basically have to keep doing whatever I start doing.

We use closed loop control in lots of situations without even thinking about it.

If a box we’re picking up is heavier than expected, we feel it pull the skin on our fingers or arms so we tighten our grip.

If it’s EVEN heavier than expected, we’d might try & involve our other hand, & if its too heavy, well, we’ll call over my open-loop-example buddy.

But this process has to be programmed when it comes to building robots.

Manipulation can look different depending on a robot’s hardware and programming.

But with enough work we can get robots to perform specific tasks like removing the cream from an oreo cookie.

Beyond building capable robots that work on their own, we also have to consider how robots interact and coordinate with other robots and even humans.

In fact, there’s a whole field of Human-Robot Interaction that studies how to have robots work with or learn from humans.

This means they have to understand our body and spoken language commands.

What’s so exciting about Robotics is that it brings together every area of AI into one machine.

And in the future, it could bring us super powers, help with disabilities, and even make the world a little more convenient by delivering snacks.

John-Green-bot: Here you go, Jabril!

Thanks, John-Green-bot… go get me a spoon.

But… we’re still a long way from household robots that can do all these things.

And when we’re building and training robots, we’re working in test spaces rather than the real world.

For instance, a LOT of work gets done on self-driving car AI, before it even gets close to a real road.

We don’t want a flawed system to accidentally hurt humans.

These test spaces for AI can be anything from warehouses, where robots can practice walking, to virtual mazes that can help an AI model learn to navigate.

In fact, some of the common virtual test spaces are programmed for human entertainment: games.

So next week, we’ll see how teaching AI to play games (even games like chess) can