Training your first humanoid with AI podcast transcript

Dale Tutt:  You mentioned earlier the software, and that I think you said 40% of the software was to deliver the robot.

I’ve got a couple of questions around the software and a couple of thoughts. But how long do companies spend on the training parts? You’ve received this robot and you want to send it out in the factory. Is this training, they load the software and it’s really good to go, or is there additional training and calibration? What are they, how long does that training take, I guess, is the question. It’s kind of the onboarding the robot. 

Rahul Garg: It’s a good question. It’s where a lot of the companies are. That’s the stage many companies are, I would say, at least the well-known, publicly known companies. And there are certain areas where production is already happening, right, using these. Generally speaking, I would say you unbox a humanoid delivered from your OEM supplier, and you could be up and running within a couple of days, right? They come pretty well set up, and you unbox it, and you do some minimal level of assembly, and you’re good to go. 

But after that assembly process, right, then you’ve got to go through that initial training, as I was mentioning. And then that can take anywhere from a week to two weeks of training, and then The thing that people are still trying to sort through, right, is carving out that work environment around where that humanoid would be functioning, right? Still having concerns of safety and whatnot, right? So how can you have that task be performed in a safe environment where people don’t accidentally go or something, at least as people are still trying to figure out the usage of these things. Typically, you can have something up and running, I would say, within two to three weeks, that’s on the higher end of anything.

But as I said, the bigger issue is just trying to make sure that you don’t create any conditions where a human is losing that same assembly line process and say, Hey, you know what? Here’s this funny-looking guy. Let me go say hi to him and give him a tap on his back or something, right?

You know, it’s interesting. I was talking to one of our customers who had introduced one of these humanoids into their production environment. And they had set up that humanoid to talk as well, right? Maybe when you would ask something or say some certain things, it will reply to you, because they are well-trained from an AI perspective, and they can answer questions that you would ask, for example, your Amazon Alexa. This is a funny story, right? So what had happened is the voice they had given to the humanoid was a nice, sweet voice, right? Everyone loved that, and people wanted to go and talk to this guy. So they said, Hey, wait a second, we need to change that. They changed it into a harsher voice. 

Dale Tutt:  Well, I mean, you talk about that, but I was at the Lowe’s store the other day and buying some stuff to work on my house, and they had one of these inventory robots, and all the kids are walking up to it trying to talk to it, and it just kind of keeps going about its job. I was like, the company ought to hire all the kids to start counting inventory too and help the robot out. But it is pretty funny, you don’t think about those interactions. 

Let’s dig into the software considerations a little bit, because we’ve talked about this a couple of times. We talked about the context of what we need to do for training and how to use AI to help train them. I was using that example and then you kind of picked up on it like, okay, now you have a parts, you have to start replacing sensors or doing maintenance on the robot. Yeah, I kind of want to dig into some of the software considerations around maintenance. 

And one of the things I was thinking about, and I’m going to use your example a little bit, but 40 or 50% of the software that’s delivered with the robot is kind of the basic functionality of the operating system, if you will. It’s the thing that helps the arms move. And then the next layer of software is the application. To use my iPhone comparison, you know, maybe I have the box installation app and I have the bolt tightening app and you have these different things that you’re training the robot how to do. And so at some point, you’re going to have a couple things happening.

One is you’ve got to change out some hardware. You’ve got to change the sensor, whatever, and you’ve got to make sure that it’s compatible. But the other thing that, they really think is going to be critical for a lot of these companies is the cybersecurity issue. When you need to load the new iOS on your iPhone because there’s a cybersecurity issue. With these robots, they’re going to be on the internet. 

They’re probably at least going to be wireless and over the air and stuff. Inevitably, there’s going to be cybersecurity issues. Companies are seeing it today even with simple automation that we are using. So, and that was always a consideration with the IoT activities, but What’s the company going to do? You got to push a new cybersecurity update or security patch, whatever, and you can’t brick all the robots. I think the considerations for how do you develop and test that software. Maybe the iPhone example is actually pretty good. I mean, you’re constantly getting updates on your phone and everything still works.

And now you don’t have to replace components sometimes. But what are these companies talking about? I know we help a lot of companies with this right now and managing that, being able to do those compatibility checks and the software checks and everything. But I’m just kind of curious. It’s like, what are the conversations having? Or is everybody just kind of so focused on figuring out how to get the humanoid robot to make the right motion that they maybe haven’t had that conversation yet? 

Rahul Garg: No, not that they have, right? Because as you rightly pointed out, the security, the cybersecurity, the software aspects are extremely important. One is obviously the safety of the humanoid work in the context of what it’s doing. But once you identify an issue, how do you update the software, right, as you rightly pointed out? And sometimes maybe it may require something more than just a simple act of training that the operator is doing down on the floor shop, on the shop floor.

There is certainly a lot of being done in that context to do the over-the-air updates, right, and putting in the necessary security elements as we spoke about. In fact, a lot of our customers are talking to us in the context of some of our applications we can help them with. to manage their entire software lifecycle and being able to do those over-the-air updates, being able to do that in a secure way, in a way that can give the guarantees, or at least gives them the ability to manage that whole process in a much more efficient way, I would say. That’s where we can certainly help companies as well, Polarion being one of our big applications that helps in that process. Customers are beginning to recognize that and beginning to start leveraging some of those capabilities. 

Dale Tutt:  Maybe this is an odd question, but you go through the training, is how much of this is like you’re training for that specific task. And that robot will keep doing that road task over and over again. But then you start to train, but when you actually get into a factory environment, the amount of autonomy that it recognizes that, hey, there’s something that’s been added in my way, and I now have to walk around it. Or how adaptable is it once it’s actually out in the factory floor? 

Rahul Garg: Some of the scenarios and the examples we have seen have actually been fantastic, right? The simple example I was talking about, the dishwasher. You can imagine the number of permutations, combinations of plates, spoons, and forks and dishes right in a small dishwasher. And there’s no way one can go train for every possible combination, but because of AI, it’s possible. We have got some of these humanoids in our own factories, and we are training those humanoids for lifting and carrying bags and totes of parts, right? 

We are not training them on every possible weight and size of what that tote may contain. and where they need to pick it up from, where they need to drop it off. Those dimensions, the weight, and all of that is changing on a daily basis, on an hourly basis, on an every use case basis. But once we have trained them a few times, then it picks up and the self learns the rest of the possible scenarios that maybe need to work in. 

Dale Tutt:  I think about some of the considerations that you have to do in order to train a robot and think about how much simulation and the digital twin really gives you the foundation to lay that foundation with the simulation and you’ve pre-programmed it to do certain things. How you’re incorporating AI then into the software and the logic of the software so that the robot is doing the right activities and making those adjustments as it needs to make them because it just can’t be on a preordained path. And so anyway, as I think about this, it’s a pretty fascinating thing to think about how do you do it? 

But I think sometimes we forget that there’s a lot of things that are happening in the environment around you that people are adjusting to subconsciously. So that same level, maybe this is where AI really starts to play a part, that is the physical AI, is the robot’s able to learn in that environment and make some of those adjustments so that this part isn’t exactly where I expected it to be when I went to go get it. And then when I go to put it in, it doesn’t fit, whatever. 

Rahul Garg: There’s one very important piece you kind of spoke about. At the end of the day, that humanoid or that robot that’s coming in is still working in a factory where there are other humans, there are other automation systems, there are other robots that are doing certain things. So there is one aspect of actually training the robot to do a certain task, but then how do you bring it in the context of the rest of the environment? And that’s where a digital twin, a comprehensive digital twin becomes even more critical, right? Where you have the ability to look at an existing plant and the line and the assembly station and down to the electrical systems and the cables that are there at that station and seeing how that humanoid will function in that existing environment.

So being able to evaluate all that in the digital twin is an absolute prerequisite. Otherwise, your system will remain in a pilot mode. A lot of companies are struggling. It’s like, how do I take this from a pilot mode to a real application mode? Whereas if you pull that into the digital twin first, then you can evaluate a lot more scenarios, a lot more situations, and start taking more concrete, incremental steps to go forward into your next step of your journey. 

Dale Tutt: You brought up a couple of really interesting points that I want to expand on and maybe ask your opinion on. Just the size of semiconductor chips have gotten smaller. I mean, you think about like the large supercomputers that used to have in a room that now has less computing power than, say, your smartphone. And that is now an enabling technology that’s part of those robots. But then the second part of it is the simulation and digital twin. and really to help train the robots. 

In the past, there was a lot of time in terms of programming and training. And I think now the tools have gotten so strong, the digital twin, the simulation tools have gotten so strong, that is actually making it a lot easier to train these robots as well. So again, it’s part of reducing that overall capital investment you have to make when you’re starting to bring them into the factory, as well as just, you know, it’s accelerating the time to deploy it. 

Rahul Garg: In the past, in the context of robots, right, it would take them a year or two or whatever, right amount of time to get it to perform a specific task. And then there would be an incremental training, which would again be programmed. The whole world of AI has changed that. The whole world of AI and simulation has changed that, right? Because what would take a year, now you can get done in a week. You give it a few. parameters, you give it a few dry runs, if you will, and then the simulation world will take over and then figure out all the other permutations and combinations. So you run it 10 times, and then the simulation world will run it 1,000 and 10,000 times.

You’re trying to get a robot to pick up something in your home, and they have literally got this going on right now. They are testing a robot at some of our customers’ facilities to unload a dishwasher, and they do that five times. And then the robot and the simulation is figuring out all the thousands of permutations, combinations of all the different cup sizes and the shapes and where you would load them, unload them, how would that all be done? And that’s all happening because of the technologies of AI and simulation that’s making it a lot more possible and the speed at which you can do this now. 

Dale Tutt:  The power of the digital twin for stimulating it, saying just the motion and how do you keep it from tipping over and you’re walking up stairs or whatever you’re doing or climbing over rocks, that having a strong foundation digital twin that allows for multi-domain optimization across electronics and software and structures really is going to be super critical. 

Rahul Garg: I would state that no company will be able to build these if they do not have that infrastructure, right? A strong digital twin infrastructure that can build all these necessary things together. And that’s where I think our whole notion of the comprehensive digital twin and also the digital threads, right? You kind of alluded to one thing, right? So say you get a supplier that’s changed the LiDAR sensors, some new version of that sensor.

Now, what do you need to do to bring that new sensor into your design engineering and eventually into your manufacturing environment? If you don’t have a digital thread that can keep all of these things in sync, forget it. You will be just in a research R&D engineering mode only, right? And the issue now is not as much about engineering, it’s more about production. 

Dale Tutt:  You’re exactly right. And actually, when you think about that, when you start to change out those parts, you have to know how that’s going to work. And if all of your robots stop working out in the factory line when they start getting updated, then that’s a big expense to the company that bought the robots. You have to be very purposeful And you have to have digital tools when you’re trying to manage all these configurations across all these different robots and different components that are feeding into it. And that’s the only way you’re ever going to scale. You’re going to actually get to the quantities that need to be produced. 

And so… Let’s talk about it. let’s transition to that manufacturing discussion. To me, it’s the same thing. You’ve got to have good simulation of your manufacturing operations. You’ve got to have good control of your operations, understanding how you’re managing your bill of materials, your manufacturing bill of materials down in your bill as a process, and just managing the quality of it and how do you do the operational tests.

As I think about that, again, you know, we talked about the digital twin in the context of the design. Let’s talk about it a little bit now and the digital twin in the context of the manufacturing. and how critical that is. I’m just kind of curious, what are you seeing? Are we going to have humanoid robots building humanoid robots or is it other robots? How do you see this coming to pass? 

Rahul Garg: For one thing I can tell you right now, robots building the humanoids is happening as we speak. That’s going on at a very rapid pace and the necessary humans are coming in for the assembly processes. There are high precision parts that need to be machined for these things. So that’s where a lot of iron machining tool capabilities and robots are coming into place as well. The tight tolerances that need to be managed, they cannot be done without having the necessary robotic actions in place, right? Otherwise you’ll not get the assembly results that you’re looking for. All that is already happening.

To the point we were talking about earlier, as companies look to figure out how to go to scale from prototype to larger scale production, and maintaining their quality and maintaining their manufacturing cost, right? One of the reasons why the adoption is slower as well, right? The cost of these humanoids are still relatively high, but they need, the cost needs to come down. And that’s where they’re all aggressively exploring different options and different ways on how to do this. And we as a company, obviously, we have a strong tradition in manufacturing, right? Strong history in manufacturing. We know how to do this all the way from robots to assembly lines to automation, to bringing in the right calibration, to bringing in the testing, the quality, the virtual commissioning of the systems. There’s a lot we can offer to a company to really help them scale up the production from prototype to large scale volumes.