Intro

00:00 Today I'm going to show you how you can train your own robotic arm to do your bidding with no code in less than 20 minutes. And yes, I know that less than 20 minutes sounds way too good to be true, but trust me, the world of physical robotics and physical AI has come so far, so quickly that it legitimately does only take 20 minutes

00:22 to get up to this point. And I'll also show you how you can train your robot arm to be able to do these tasks autonomously as well. That task does take a little bit longer than 20 minutes, but it's still being able to be done all under $300 as well. And it's also able to be able to correct itself from a mistake from a previous position. Oh, and there we go. So, it made a

00:45 mistake. It was able to pick itself back up. And I'll show you how you can do all that with no code. Now, I'm Broen Pratt. I've been teaching robotics and design professionally for over a decade now. So, let's skip some of that fluff and let's get into how are we actually going to do that learning to be able to make the robot do that. We're going to run this in about six steps. We're going to talk about what the ARM is, how do you set up your computer, how do you set up the arms, train them, train the eye

01:08 model, and then actually run the model eventually, but let's get a little deeper into what each of those parts actually mean. In part one, we'll talk about what is the SO101 arm from Low Robot, and we'll talk about the parts you need. You can get this up and running for under $300 US. I'll show you how you can either DIY 3D print it yourself or how you can get a preassembled arm um to be able to get

01:30 this up and running even faster. We'll talk about the sort of computer specs that you're going to need as well as the sort of ideal uh imitation learning environment you're going to need to be able to teach your robot how to do that. So, I'll I'll tell you all about this information so you can have really good data. Then, we'll talk about the actual computer setup. I'll teach you how to install the software that we're going to need to be able to run this. This is

01:53 using something called the lab. Then we have to get into things that are arm calibration and teleyup. So when you first connect your robot arm, all the motors don't know where they are in space. They don't know how far they can move. And just like if I were to take my shoulder and move it up much further past this point, I might break a joint. The same thing can happen to your robot. So we'll teach it where its safe limits

02:16 are. We'll teach it where these things are connected to your computer, how to set up our cameras, and then we'll actually get into teley operation as well. And teley operation is that when you move your leader arm, your follower arm does exactly what the leader was doing. And that part's going to take you upwards of about 20 to 30 minutes to be able to get up and running. Then the part that's actually going to take a

02:37 little bit longer is actually training that arm to complete a task. The AI model we're going to use is something called the ACT model or the action chunking with transformers. I'll talk about what that is and how does it actually work to teach your robot to be able to follow along steps. We'll do a test recording for our data so we can have our test if our test was to be able to pick up this pen. We'll do a test recording to show the robot how to do

03:00 that. I'll show you how to manually replay those data episodes to watch it in real life. And then we'll actually record a full set of training data. Then we can take that full training data and actually feed that into our AI model, that ACT model to actually have it trained for us. I'll show you how if you have a powerful computer, something that's got a Nvidia GPU on it, how you can train it locally yourself for free,

03:22 besides the cost of, of course, the graphics card and the energy you need to be able to run it. But if you're like me and you don't have a very powerful computer, I'll also show you how you can use some online software using Hugging Face Jobs or Google Collab. And for under a dollar, it cost me 72 cents to train the model you saw that was able to pick up that little orange dinosaur to

03:46 be able to get that up. We'll talk about actually train that policy and then what does it actually mean to train? What is that data and how does training actually work? And once we've evaluated that uh training, we'll take a look at what the model actually looks like and we'll show you how to actually run it in real life so that your robot can actually run that task independently. And not only is it going to be able to just pick up those

04:08 individual objects, it's also going to be able to pick up those objects and recover from a fault or a failure as well, which is really slick about that. And along the way, I'll give you lots of debugging notes, tips, and tricks. Now, when it comes to the parts you're actually going to need, you got one of two options. Uh the first option is you can take a more DIY route. If you have access to a 3D printer, that could be a