EP 485: Humanoids in our world. How it’ll work and what’s next
21:16 Share

This is the Everyday AI Show, the everyday podcast where we simplify AI and bring its power to your fingertips. Listen daily for practical advice to boost your career, business, and everyday life.

One thing that's happening here at NVIDIA GTC, there's much more than, you know, large language models and GPUs. One exciting area of AI, it's humanoids, it's robotics, it's embodied AI, how we can take all of these innovations and make our actual, uh,

Worlds that we live in better, our jobs hopefully safer, and maybe even more enjoyable. So that's one of the things that we're going to be talking about today on Everyday AI. What's going on, y'all? My name is Jordan Wilson, and I'm the host of Everyday AI, and this is a daily live stream.

podcasts and free daily newsletter helping everyday people like you and me not just keep up with what's happening in the world of AI because there's a lot but how we can use it to get ahead to grow our companies and our careers. If that sounds like what you're trying to do, you're in the right place and I'm excited for today's conversation because one thing you don't have to be scared.

of humanoids, right? We're going to understand in today's conversation what they actually are and talk about what is the future of humanoids and robotics and how they're going to impact the future of our work. All right, but don't worry if you're on the podcast, maybe I sound a little different, but I'm actually reporting here live at the NVIDIA GTC conference where we're extremely lucky to be able to talk to some of the

leaders bringing AI, in this case, to the real world. So please help me welcoming to our livestream audience, at least, we have Pras Velagapudi, who is the CTO of Agility Robotics. Pras, thank you so much for joining the Everyday AI Show. Thanks for having me. All right. So before we get into, you know, humanoids and robotics and talking about all this, first, tell us a little bit about what it is that you all do at Agility.

So at Agility Robotics, we have our humanoid robot, Digit. It's a logistics and manufacturing focused robot right now, but it is a humanoid with the ability to move and work in human spaces. And so we can do a lot of tasks like moving around bins and loading, unloading equipment in a form factor that doesn't really require you to modify any of the spaces that you have already. You can use human shelves and human totes and other things like that. And so Digit

what we're building out is basically this platform to be able to take our robot out into the human world and be able to do all sorts of work, starting in logistics and manufacturing, but ideally moving into retail and maybe one day the home.

Okay, I'm excited to talk about that, even getting humanoids into the home. But I want to rewind a little bit, right? Because even at this conference last year, right, there was a lot of talk and excitement around humanoids, right? Can you talk a little bit, bring us to the current day? Where's the space at?

right now not just the the amazing work you all are doing in agility but where's the humanoids uh you know kind of uh progress at because it seems like it's it's always changing it's hard to keep up with what are they capable of are they actually out there you know doing jobs today like where are we at

Yeah. So it's been moving really quickly. You're definitely right about that. In the past year, what we've seen is this convergence of technologies, both in the hardware, in things like energy storage and actuation, and in the software with all of the advancements in AI has really enabled a lot of technology to come together to make the humanoid platforms that you see in a lot of videos and things like that today.

And we're right at the cusp of this kind of explosion of the capabilities of these platforms being able to make it out into the world.

And so where we are right now, I think, is that there's a lot of humanoids that are emerging in the market, and there's a few that are making the transition to being able to do true useful work out there. So we're one of them. We have customers right now. Our robots can work in full shift operation in facilities, and they do, where they're working full eight-hour days. So we're just starting to see that happening in the market where

there's actually this ability for a humanoid robot to come into something like a logistics or warehousing or manufacturing facility and do a job, a useful job. And usually the types of jobs that really it would be great for humans to never have to do moving around heavy objects, crouching really low, loading and unloading things. These types of repetitive tasks that

Really, we have humans doing it mostly because they're inconvenient to automate, not because there are such valued, loved jobs by the people doing them. All right. So I have to have a follow up on that right away. Why are the humanoids only working eight hours? Why aren't they working around the clock? Right. So great question. They absolutely can work longer than that. It's just we happen to be in some facilities right now that that's when the rest of the system is working because there are humans upstream and downstream of us.

But the robots themselves, you're absolutely right. We can run two or three shifts with the same robots. They don't care. They can run continuously. But we are limited. We're not limited. I'd say we have to be matched up to what the process around us is doing, which is sometimes robots on either side of us doing other types of tasks and sometimes humans upstream or downstream that are doing other tasks that are related to the overall flow.

So I do want to get into what's new at GTC. But first, I want to talk a little bit about your involvement in the Inception program. You know, a lot of great and promising startups. But can you just talk a little bit about what your experience has been like in the NVIDIA Inception program?

Yeah, so it's been a really great connector for us. NVIDIA in general has been a really great partner for Agility and helping it out in a lot of different ways because we both use NVIDIA hardware, we use some of their software, and we're really

really aligned with their product teams in terms of figuring out what to do next and how to focus some of the new technology that's coming out. And so the Inception program has been a great connection point for that in being able to get us training, get us access to some resources that we can use to help accelerate our adoption of NVIDIA technology.

and really just get us connected and supported in using the pieces that NVIDIA has for us. So speaking of NVIDIA technologies, a lot of new announcements this week at GTC. Let's talk about what's new and what are you excited about for Agility?

- Yeah, so we're showing off a demo here at GTC where we're using NVIDIA Isaac Lab trained policies to do whole body motion control of our robot. So basically between last year's announcement and this year, we've actually adopted a lot of that technology and gotten it working and now have a control stack that's gonna be picking and placing retail grocery items

using a fully AI-trained stack that went directly from NVIDIA's simulation environment to the real robot with no other data, which is really an exciting step for us. And so we're particularly proud of that. And we're continuing to build out this sort of Isaac Lab ecosystem for ourselves.

We're also working on NVIDIA Mega, which is a platform that NVIDIA announced, I think, back at CES, which is intended to basically support this distributed workload and simulation of, for example, multiple robots working at something larger, like a facility scale. So we have a customer, Scheffler, that we've been working with that we're essentially a

building out the pieces for them to be able to develop out larger scale simulations of things like their entire facility where they might be using digit robots and parts of their flows. And so we're building up the pieces to be able to do that. So mega overall is sort of lifting up another level in the robotic space into not just thinking about how do you train and run an individual robot, but how do you train and run

fleets of robots across a facility. And so it's a pretty new tool and we're excited to see where that leads. Yeah. And let's get even more elementary here. Is there a difference between a robot and a humanoid? Is it the same thing? Is it just as the capabilities and the technology gets better, we just refer to them as humanoids as they take on more tasks that maybe demand more cognitive function? Is there a difference between a robot and a humanoid? Yeah. So...

Great question. And I think probably over the years, this sort of terminology has sort of evolved, right? Based on our conception of what robots are capable of. Humanoids are, they're a class of robots. They're typically used to refer to robots that either look a lot like humans in terms of their form factor or can do things in human environments to some extent. Like they can operate in a human home, in a human space and live

do things the way that humans would do without special accommodations. So I think it's maybe a little bit different from saying that all robots will eventually converge to humanoids. That's probably not going to happen. There's a lot of very effective robots that are good at what they do in other form factors other than humanoid form factors.

especially when you're talking about things like transporting objects around or dealing with industrial processes where there's very specialized equipment or needs that robots designed for that function can do very effectively. Where humanoid robots can really shine is in being human-centric.

in not having to change their environment in order to do their tasks. A humanoid robot can use the same types of containers that you would carry around. In fact, we are in the demo that we're presenting at GTC, we're using a shopping basket. We just bought

off of the internet from a place that sells shopping baskets. There's no special accommodations, right? We're using a shelf that's just literally a store retail shelf, right? And when we're in our customer facilities, we're putting the robot into flows that were previously ones where human labor was doing the work of lifting and moving stuff around. And so the power of a humanoid platform

I think is less about, okay, it's specifically got two arms and two legs and is about so-and-so high. It's more that, well, I can move into the same spaces that you do. I can use the same types of items that you do. I can do the same types of tasks that you do so that I can come into your environment without you having to restructure everything, often at great expense to instrument it, reorganize it,

and tool it up specifically for any particular robotic piece of things. - So it seems like, you know, 'cause I was here at GTC last year and I, you know, remember Jensen coming out and, you know, talking about Isaac and having all the robots, but you know, how has the humanoid space kind of evolved even over the past year? I mean, is it very common to walk into a big, you know, logistics or warehouse and seeing humanoids or are we still not quite there? Maybe that's what's coming in 2025.

But I would say we're on the path there versus where we were last year. You can at the very least go into some warehouses and see some humanoids, which is definitely different from even a year or two ago. Right. We have we have a multi-year contract in place with GXO, for example, where there's humanoids working in in a facility in one of their facilities all day, every day.

I think adoption is not quite at the level where that's at every warehouse, but now it's definitely been established that this is a thing that's possible, that you can get real value out of doing it. And we're also seeing a great acceleration in the capabilities and the

speed at which we're seeing evolution in the platforms that are available. So the performance of the systems is going up, the types of capabilities and flows that we're able to take on and sort of like get too close to human, not

not at human performance, because you don't necessarily need that, but enough to be valuable to someone to not have a human doing the same tasks, we're able to cover an increasing amount of that space. So I'd say that for us, we're really seeing this ramp in velocity. I think if you look out into the media space, you're seeing a lot of really cool demos of that functionality in lab environments right now and

I think that's going to, that inertia is going to continue into what we'll be capable of in the real world and reliable in industrial settings and manufacturing settings and things like that to start out and then move from there. Yeah. And where do we move from there? Right. So whether you want to talk specifically, you know, agility and the type of, you know, companies that you're looking to work with in the future, or just more generally about the types of work, but, you know, you know, manufacturing warehouses,

Does that make sense? But where might we be going next? I mean, I'm assuming we're not going to have humanoids having like desk jobs, right? Traditional desk jobs, but you said they're probably going to end up in our homes, but what's another type of work that humanoids might be very well suited for outside of manufacturing? - Yeah, so first of all, there's plenty of manufacturing to be done. Logistics and manufacturing alone is tens of thousands, hundreds of thousands, maybe millions of robots right in that space alone.

but going beyond that you can think about things like working in the back of the store and retail um things like restocking shelves things like moving uh material around in uh hospitals or other types of environments and when you think about okay why these environments and not other ones why are we even starting in logistics and manufacturing there's actually a pretty good reason which is that those are the environments in which the structure and the um

training are most amenable to meet humanoids where they are in terms of safety.

So a humanoid robot, especially one that can do useful work, it's got a lot of capability, but that's also a lot of energy and force that it can use to do things in the world. And unlike other types of robots that just need to avoid ever touching anything, which a self-driving car for most of its lifetime is mostly concerned with not touching anything. That's its measure of success.

You get in the car and then it gets to the destination while not touching anything else. But with the humanoid robot, a core part of what it's doing is touching stuff all the time. And that means that we really need to understand, okay, how do we safely impart our forces on the world?

Logistics and manufacturing has a long history of using automation. And so it means that there's a good starting point. The rules in some sense are more understood. As we expand out from there, we kind of have to figure out what those rules are going to be in other parts of society. If we were to put a humanoid in the home like today, well, there's a lot of gray area in terms of exactly how

how you would ensure that it could be safe and what types of things around it might be reasonable. How does it handle things like pets or children? Is it okay for it to be carrying a hot pot of something? Because even if it doesn't cause a problem, it could spill something. All of those are, I think, societal things that will take some time to

be figured out, right, both our comfort levels as a society and also how the technology can advance to be able to provide better guarantees about that stuff.

but one place where we can do it right now is in logistics and manufacturing and then from there to things like commercial applications. So that's why that evolution I think is the likely progression is because it follows basically where we have a better idea of not just how to make the humanoids do the work, but also how to safely get them out in the world such that when they're being relied upon to work, you know,

every day, all day, that all those statistical edge cases of like, what if you have a slippery floor one day? Or what if somebody miss packs something it's overflowing? Like, will you like tip it over and things like that? All those can be reasoned about and covered. - Yeah.

I think even when in a lot of probably our audience, their day-to-day right now interactions with AI are using large language models and fine-tuning them on their company's data and bringing in rag pipelines and all of those things. But when business leaders who are listening now and they're

maybe very curious about how they might be able to integrate, you know, humanoids into their workflows. There's probably a bit of, you know, maybe some, some, some apprehension, right? Because I think with AI, it's like, okay, well, I'm going to go in, I'm going to tell a chatbot this, you know, I'm going to initiate something where humanoids are essentially, you know, they're out there, they're kind of doing their own thing, right? And that's what they're programmed to do. So, you know, how can you, you know, what are you all doing to kind of address

the whole safety piece? Because I know that there's, you know, sometimes people, you know, think, oh, you know, this is just Terminator, but it's not, right? So like, how do you address the safety and the guardrails of a humanoid? So right now, what we do is by having a completely independent supervisory system. So we have our safety system and we have our control system and they sort of operate in parallel.

And that's the easiest way to do it. And so it's a good starting point. But right now we tie that external safety system to whatever the robot's operating within. So that might be a work cell. It might be something like laser curtains or external sensing. We basically pair the robot off with some aspect of its environment that's used to tell, you know, how close are humans getting and where could hazards be introduced. Now, where we're going from this is to take all of that information

sensing and reasoning about where people are and where we could induce these hazards and bring it on board the robot. So that's what we're doing over the next year is basically building out an on-board safety system on the robot to be able to get to what we call cooperative safety, which is humans and robots being able to safely be in the same space. And we want to achieve that without requiring anything special in the environment the way that we do right now. So I think that is kind of

how we can get to this kind of safe operation without requiring any sort of unobtainium, like we require generalized AGI. It's like, no, no, we just require a very well-designed safety semantic about like how the robot responds to people that's run on a reliable, verifiable system. And then we kind of run it in parallel to the AI models that are making decisions about performance,

like how to move quickly, how to grab things. We're sort of running a separate parallel system, which is just reasoning about is the thing I'm doing going to cause a hazard or not? So we've covered a lot in our short conversation already, but as we wrap up, what do you think is the most important part

or maybe even most exciting takeaway for you and what Agility is doing in announcements here at GTC. What do you think is going to be, you know, maybe that thing that is going to be most impactful for the everyday person and how they work in the future?

I think the biggest thing is that we're seeing that humanoids are a real thing and they're here to stay, right? This is now just a new piece of the puzzle in robotics and automation. It's not some far off abstract concept or a thing that's in a lab. It's,

It's okay now when I want to choose how to do something in the real world, one of the options is just a humanoid robot and the performance and the capabilities are only going to get better from where they are today. And they're doing so at this just astonishing pace.

So I think people who are interested, you know, check in, take a look at some of the stuff that we can do, take a look at what the space can offer. And basically you should keep checking in because I think every six months, every nine months, that waterline is going to keep going up at an astonishing rate. All right.

Right. It's extremely exciting to watch and follow this space. And, you know, hey, you know, he mentioned a lot about some of these demos. So we're going to be, you know, sharing those in our newsletter. So make sure if you haven't already, please go to youreverydayai.com, sign up for the free daily newsletter. We're going to be

recapping today's conversation. And for the podcast audience, you'll be able to see a lot of what Pross was just talking about in action. So Pross, thank you so much for taking time out of your day to join the Everyday AI Show. We really appreciate it. Thanks again. All right. Thank you so much for tuning in. A lot more exclusive insights, talking with some of the brightest minds in AI at NVIDIA GDC. Thank you for tuning in. Hope to see you back tomorrow and every day for more Everyday AI. Thanks, y'all.

And that's a wrap for today's edition of Everyday AI. Thanks for joining us. If you enjoyed this episode, please subscribe and leave us a rating. It helps keep us going. For a little more AI magic, visit youreverydayai.com and sign up to our daily newsletter so you don't get left behind. Go break some barriers and we'll see you next time.