Humanoids: a new step in automation evolution -podcast transcript

Dale Tutt: Welcome to the Industry Forward Podcast, the show where we examine industry and technology trends with the help of various experts from around Siemens and beyond. My name is Dale Tutt, and I am the Group Vice President of Industries at Siemens Digital Industry Software, and I’m also your host for today’s podcast. In a recent episode, Rahul Garg and I had an exciting discussion about the major shift towards humanoid robots that’s happening in many industries. Rahul is the global head of the industrial machinery team here at Siemens, and I am welcoming him back for what I hope to be another exciting episode on the convergence of technologies that is enabling the growth of the humanoid industry, as well as the many applications of humanoids. 

So Rahul, one of the things that I am, I guess, curious about is why now? What’s happening right now? Why are people so quick to move towards adopting humanoid robots? This is a concept that’s been around for a long time. I wonder what’s driving it. We have a lack of a skilled workforce, and so maybe there’s more demand from factories to try to figure out how to use humanoid robots. to fill jobs that in the past has only been able to be really filled by an actual person working on this. You have to be able to make decisions. You have to be able to reach into certain spaces, whatever. There’s a little bit of a lack of skilled workforce. There’s situations, I think, where you maybe would be safer if you can actually take the person out. Think about mines or you think about working in a chemicals factory or other things where you require a humanoid form to work, but you can maybe improve the safety Maybe the ergonomics, a lot of repetitive motion. 

Maybe there’s more demand for it now. But I think now it’s also the convergence of technologies that have now come about with semiconductors and more software definition, physical AI. The battery technology has now gotten good enough that you can let these things work on their own. They don’t have to be dragging around cords and have long tails on them. I think it’s like two things. One is there’s a need for workforce. And two, there’s also just that the convergence of technology has actually happened to help drive this. So what are your thoughts about this? Why now, and what’s happened that’s really starting to drive this? 

Rahul Garg:  You summarized it well, Dale. It’s the convergence of technology, especially the AI and the physical AI that’s coming into picture. That was not there a few years ago. So the technology has really come together all the way from the chips, the hardware, to the batteries, getting all of that compressed enough. A lot of it is coming from the automotive industry, leveraging the things around autonomous driving, for example. But these core technologies around AI, around machine learning, around computer vision, that is one big factor. Number one, that’s one of the biggest reasons why there’s a strong ability to get these moving a lot more faster. If you think about this whole context of generative AI, it did not exist three, four years ago at that same level. To me, I think that’s the one big driving factor, the ability to do reinforcement learning, the ability to build these foundation models. All that has dramatically changed that process. Between what I saw at CES to now, just a couple of days ago, I saw recently released a video of Boston Dynamics’ latest version of Atlas. The fluidity with which it’s making the motions is unbelievable, which did not exist a month ago at CES. They’re able to do that because of the training that they were able to provide. That whole speed of how training and AI is being used is one big factor. And then the other things around the compute power, right? Being able to have that compute power delivered and on the edge, on the device. We’re bringing in all of the necessary GPUs that were not available in the past. The battery capacities, there was a big challenge with that. That’s improved tremendously. They can get them down to a form factor that’s small enough that you are not constrained with weight, gravity, stability issues, right? That can be managed a lot more easily. That was a big issue before. Where do you put the battery back if this thing has to move? That’s been solved. The sensors have improved dramatically, right? The vision sensors, the LIDARs, the cameras, the tactile sensors that can touch and feel. Some of that technology was not there as much even two, three, four years ago. That’s another big convergence that’s dramatically helping.

And then the whole interdiscipline mechatronic design process, bringing in mechanical systems with the actuators, bringing in the torque control system, the gear reduction issues, managing that with the necessary electrical systems of being able to power these motors in a efficient manner, right? So, a lot of those capabilities have all finally come together in the last couple of years. And the research curve has shifted in a big way. It’s really taken off a hockey stick curve from a research engineering side to the point like, hey, you know what? We can very easily get them onto a shop floor now. The second is the labor force, the aging workforce that’s increasing demand for these things. And then the third vector, I would say, is the cost and the capital investments, right? Costs are collapsing significantly. The cost of GPUs is coming the cost of batteries are coming down, the cost of the processing power that was needed for the AIs is coming down.

And the capital investments that are from the governments to the private sector, that’s enabling a lot more faster adoption of these capabilities, right, a lot faster research. So to me, those are the three driving factors that are making this possible. And in addition, the things that are making it possible is because of production scaling, right? So people are already going from prototypes to factories. Tesla has already committed to making 100,000 of these by 2026. Agility is talking about 10,000 of these. Many of these Chinese companies have got 5,000, 10,000, 20,000 already planned and actually already produced. What used to be a playfield idea, concept idea, now has actually moved from prototypes to the actual factories. And we have got customers who are using it, right? They’re only talking about publicly. Amazon is talking about it. BMW is talking about it. BYD is talking about already bringing these into their production operations for initial commercial scaling, right? So those are some of the bigger reasons, I would say, why we are going to continue to see more explosive growth in the segment. 

Dale Tutt: Why do you think that there’s so much emphasis on having them look humanoid and have that level of functionality versus maybe a little simpler or less humanoid looking. You might be able to question why you would have legs and something looks like a head. It really does take on the shape of a humanoid. What do you see that advantages the shape there? 

Rahul Garg:  Good question. One of the biggest value propositions and the reason why there’s such a big focus on this form factor is because it enables a very easy drop-in. Especially in brownfield environments. If you think about it, our world that we live in over the last three, four centuries, it’s been designed and engineered for a man-made world, right? For us as humans to function in it. Everything around us, our homes, our factories, the way we do things, it’s all around our form factor. So that’s what the humanoid is looking to do, is basically have the ability to fit into that infrastructure without significant changes and the ability to interoperate with the human. 

Dale Tutt: So we’re seeing a lot of applications in factory environments, but I wonder about using humanoid robots in less safe environments, fields that tend to be traditionally, you know, a little bit more dangerous, whining, working around hazardous chemicals, firefighting. Where’s the focus now and where do you think it’s going from there? 

Rahul Garg:  The way I’m beginning to see is there are two things that are going on. One is some special purpose applications being identified and designing the humanoid to address that specific need. And then the second is the more general purpose humanoid being able to set up to do many different tasks. In the context of the special purpose task, we are already beginning to see humanoids and some form factor that’s derived from a humanoid designed to do a special task. Just the other day, we were looking at an application for underwater welding. That’s a highly skilled capability, and trying to find welders who can go underwater and do the necessary task, it’s not easy. So, you know, taking on something like that is very valuable for anyone. 

Doing a lot of defense, safety-related things which would have not been easily possible without some kind of an automated thing. Just in fact, another example, we were talking to a company in Canada that is building these humanoid robots for paint shops and for low volume, right? Basically, a paint shop and an automotive OEM, those are well designed. Same continuous job you do day in, day out for hours together. Whereas if you go to a job shop, they change the car every day, two hours every hour, right? It’s a very dirty job. So they have come up with a solution for that. Those are what I would call as the special purpose robots, humanoids that are being designed, and they’re being further trained, tailored for that specific application.

 And then we have got the general purpose ones, which are being designed for a whole slew of activities and that are being put into various production facilities, at least at the moment, right? Today, at the moment, a lot of these general purpose ones are being designed for, or at least being built for the manufacturing world, for the industrial manufacturing world. And then the eventual goal is that one of these days we may find one of them in our homes, helping fold our laundry or do some other things. Just actually on that example of helping fold our laundry, there is a company that’s built a special purpose industrial robot for folding of laundry. You’ll be amazed to know that how tedious a task it is when you have to do this day in, day out for 10, 12 hours. when you are dealing with laundry coming in from hotels, hospitals, and you name it, right? Industrial applications, businesses. They needed someone who will be able to fold the laundry, the towels, or whatever. There are these mega laundromats, if you will, right, industrial laundromats, and they’ll become a perfect use case for something like this. 

Dale Tutt: Pretty fascinating, because a lot of the applications right now that are getting a lot of visibility are the large-scale usage of robots, maybe in factories and stuff, but thinking about having robots that are folding laundry, very repetitive tasks. It makes a lot of sense. I don’t know if having a robot in the house is going to make us even softer. One thing I want to dig into a little bit that you started to talk about earlier was some of the advantages of the humanoid robots. They’re being capable, I forget exactly how you said it, but they’re capable of doing more different tasks and you can task them differently. But how much flexibility do you think, like if you put a humanoid robot in a factory, to have them programmed for different things, or are they kind of just self-learning? I have this vision that you have to train a humanoid robot to do the different tasks. If you’re installing equipment on an assembly line and you have to be able to install the right equipment in the right place, and some vehicles, especially like if you have cars, they have different options. And so all of that’s got to be recognized, and the humanoid robot needs to know how to do all that work. Just dig in a little bit about that learning piece. 

Rahul Garg:  For sure. Today, there is the whole perception context that needs to be created for that humanoid to function in that environment is obviously very important, right? At the end of the day, otherwise it’s a bunch of wires, motors, and power, energy, battery packs, right? So the intelligence that needs to be built into these systems, if you will, is extremely important. And to build that intelligence, there is tremendous amount of training that needs to be incorporated. So today, the way it’s being done, or at least at today’s age, and as this field evolves, it’ll probably mature a lot more. When a humanoid does come out of a factory, it’s 40, 50% trained to kind of, you know, go around with the certain functions, right? So it can walk stably, it can move its arms, legs, move around from place to place, detect objects in its way, kind of turn around and do certain motions as necessary. So that’s like the baseline that they are calibrated with. Once that’s done, then they get to the next level, which is trying to figure out, okay, what’s the general task we need them to perform?

Is it picking up some bags or totes from one place to the other? So generally speaking, we need them to have dexterity in their arms to lift something and kind of go from one location to the other. Or is it going to be designed to, as we were talking about, doing some welding or something, right? So that becomes the next level of calibration work that needs to be done. So typically, that becomes another 30, 40% of the work from an automation perspective. And then the last 20% is actually for that specific task that they would need to be performing. And that’s where a lot of the work is also being done, is how to get that training programming done in the most easiest way. That’s when it becomes more job-specific, and the more you can simplify that process, the easier it becomes, right, from an adoption perspective. That certainly is an important element in trying to figure out now that you’ve trained the robot to pick up that tote box, that’s a high level, but now we needed to pick it up from this part to this part for this job number, for this task. So that piece of the next level of fidelity needs to be trained as well. So that’s where a lot of the work is going on is how do we simplify that piece of that training? 

Dale Tutt: Wow, thank you, Rahul. That is great insights on the training for the robots, and I appreciated that. Unfortunately, it is time for us to wrap up. So, Rahul, thank you again for sharing your insights and perspectives on such a hot topic, and I’m sure our listeners appreciated the discussion, and I hope they enjoyed it. It’s exciting for me to see the convergence of the technologies that’s coming together, semiconductors, miniaturized electronics, software, battery technology, that’s really enabling the growth of the humanoid robot industry. And since these humanoid robots are designed to operate in a world designed around humans, they can easily be used in hazardous environments or to take on highly repetitive tasks that often lead to injuries, which I really thought that was a great part of the discussion. And when we think about the commercialization with the improved simulation modeling with the digital twin, we see training and production costs starting to trend down, which will, I think, in itself lead to larger scale commercialization. So, you know, as we said in the last episode, humanoids are not a silver bullet, but they clearly have a role to play in multiple industries. And they’re moving into real production use, especially for companies that are already on their own digital journey. So Rahul, thanks again for sharing your insights today. Thank you to everyone for listening. If you enjoyed the show, please subscribe to the Industry Forward podcast on your favorite podcast provider. We’ll see you next time. We’re going to have some further discussions on humanoid robots and the technologies that’s helping them come about. So thank you. We’ll see you soon.