The Ultimate Guide to Robot Localization Methods – Transcript

00:00How does a robot answer the question, where am I? Today, we're going to find out. >> And then the robot's going pixel. So inside the pixel, >> every mobile robot from a simple toy to a self-driving car needs to answer this question. And its answer is a process called localization. I'm Broen Pratt and I've been teaching robotics and design for over a decade. And today I'm going to break down what localization is, how it can differ, and I'll explain the most common methods that robot use to solve this problem. from simply counting the

00:26number of times a wheel is rotated to more complex methods like using lasers. So, let's start off with the big idea of localization. Localization is simply the process that a robot uses to figure out two things, its position and its orientation in the environment. This process can be broken into a relative position and absolute positioning. In relative positioning, a robot's trying to figure out where it is in relation to where it started. Think of it like giving directions by saying, "Walk 20 steps forward and then turn left." You

00:53don't know where you are in a city or a country, but you do know where you are relative to your starting point. In absolute positioning, a robot's trying to know its exact coordinates on a pre-existing map. It's kind of like finding your location on a blue dot on Google or Apple Maps. You know your precise global or local coordinates.

01:12Now, I'm going to let you in a bit of a trade secret. Most robots never really rely on one method for localization. Typically, they use a technique called sensor fusion. And this means that they combine data from multiple sensors, be able to get a much more accurate and reliable result than any single sensor could provide on its own. It's kind of like how you use your eyes and your ears and your sense of balance together to be able to walk around. A robot does the same thing. So for localization methods, let's look at the simplest ones. And generally, this is used for relative

01:40positioning, and it's something called wheel odometry. And the concept is pretty straightforward. You can attach a sensor, it's called an encoder, to a robot's wheel, and a robot can count precisely how many times that wheel has spun. If it knows the wheel circumference, it can then calculate how far it has traveled. Imagine you got a basic warehouse robot whose only job is to pick up a bin, drive 10 mirrors down in a perfectly straight and clean aisle, and drop it off. For that short, simple, repeatable task, odometry is cheap. It's

02:08really easy to implement, and it might be all you need for your robot. The problem with using odometry over wheels that in really long distances or more complex environments, some errors get built up in the measurements of those wheels rotation. Wheels sometimes slip.

02:22The floor might have bumps and this can cause the robot to calculate position to drift away from the ax position where it is. It becomes unreliable over long periods of time or in unstable environments. This is where our next sensor comes in, the IMU or inertial measurement unit. Uh your phone has got one inside of it. It's a chip with an accelerometer which measures changes in speed and a gyroscope which measures changes in rotations. You can think of a

02:48drone. An IMU is essential for flight stability. If a gust of wind suddenly tilts the drone, the gyroscope detects this rotation and tells the flight controller to adjust its motor speed to be able to counteract that shift and it can keep the drone level. The IMU doesn't know where the drone is in the field, but it knows its orientation has changed from where it was previously and it can react accordingly. While relative positioning for localization can be pretty helpful, doesn't really get us close to what robots are really good at, which is doing repeatable tasks. And

03:16this is where something called a visual fidial marker, like an April tag or other QR code style codes, can really help. Before the robot even starts up, you give it a map. And this map tells the robot its precise coordinates of every single April tag. For example, tag number 27's at position X5 and Y10 inside this room. Then, as a robot drives around, its camera is constantly looking for these tags. And when it sees tag 27, it knows the tag's exact location from the map. It measures its

03:44own distance and angle to the tag. And from that, it can calculate its own absolute position in the room with some incredible precision. This immediately corrects any drift that might have accumulated or that error from those odometry wheels. A fantastic real world example of this is inside an Amazon warehouse. The robots navigate by reading markers on the floor, allowing to find specific shelves with perfect accuracy or near perfect. But what if a robot can't put up markers everywhere?

04:11And what if they're an unknown environment like your home or disaster relief? For that, robots use something called SLAM, which is simultaneous localization and mapping. Slam is an algorithm that solves a classic chicken the egg problem. To know where you are in a map, you need to have map. But to build a map, you need to know where you are as you move around. So that slam algorithms cleverly do both at the same time. You can the best analogy is kind of think about like you're in a dark

04:38room. You'd use your hands to feel around for walls and furniture. You build up a mental map of the space while also keeping track of your own steps, your movements. A robot does this with a camera or more commonly with some LAR sensors, which is a spinning laser that you kind of see on robot vacuums. Locus Robotics is a great example of a SLAM algorithm in action. It's a pick and place a robot and it uses one of these methods of slam. The first time it runs, it uses SLAM created floor plan of the warehouse. On every future run, it uses

05:05that saved map to know exactly where it is, where it's been, and plan the most efficient route through the map. If we need to know localization in a huge space, we have a method that we all know, GPS, or the globe positioning system. For robots that need to operate over large outdoor areas, GPS is a great go-to for absolute positioning. But the GPS in most devices, like your phone, is only good to a few meters, which isn't really good enough for what most robots need. If we look at modern farm tractors, they're equipped with a

05:31special type called RTK GPS. They can get to centimeter level accuracy. This allows them to autonomously plant seeds, spray fertilizer, and harvest crops in almost perfect straight lines year after year after year, maximizing land use, and minimizing some waste. It's wild stuff. The most advanced robots of today, like a self-driving car or human robot, are great examples of that sensor fusion in action. They may use wheelometry, multiple IMUs, cameras, and LAR for SLAM, or even GPS. And they're

05:59constantly blending all this data together to be able to create a single, highly accurate and incredibly reliable understanding of their place in the world. Typically, one data point isn't good enough. I'm curious what other examples of robots you've seen in real life that have different localization methods. Let me know in the comments down below. And if you'd like to access more robotics tutorials, CAD, co- snippets, or more, you can consider joining the community. If you have any questions, drop them in the comments, and I'd love to hear what it is you're

06:26working on. And as always, best of luck on your next robotics