[00:00:00] Pras Velagapudi: We do paid labor. We have customers that deploy our robots in their manufacturing facilities and their logistics facilities and Digit works in a full shift and gets paid for it. So that's the real value of creating a human centric robot like Digit. It's to be able to be flexible and easy to deploy, and all of these different places in a in a process flow that were traditionally very difficult and specific. I would say if you are a humanoid robotics provider yourself, I would encourage you to understand safety and start, you know, working with the standards bodies that are that are working on that.

[00:00:42] Audrow Nash: Hi Pras, would you introduce yourself?

[00:00:44] Pras Velagapudi: Hey, I'm Pras Velagapudi. I'm the chief technology officer at Agility Robotics and at Agility we are working on building Digit, the humanoid that's made for work.

[00:00:55] Audrow Nash: Hell yeah. Tell me about Digit and tell me how it's made for work.

[00:00:58] Pras Velagapudi: Yeah. So Digit is our human centric robotic platform. So it's a robot with two arms and two legs. It's spun out of technology coming out of Oregon State University. And a bunch of related academic research around how to create. A dynamically stable robot. And that evolved into the platform that we created, which is doing right now material handling. So moving around totes and bins and things like that in warehouses and logistics and manufacturing applications. And so we're building up a robot that's really intended not just to be a thing that looks kind of human like, but something that actually can go into environments that are designed for humans, those parts of facilities, and do useful work. We do paid labor. We have customers that deploy our robots in their manufacturing facilities, in their logistics facilities, and it works in a full shift and gets paid for it.

[00:01:52] Audrow Nash: Awesome. And yeah, you just celery celebrated, an anniversary of Digit working somewhere for a while. Would you tell me a bit about that?

[00:01:59] Pras Velagapudi: Yeah. So we recently celebrated our one year anniversary being deployed at GXO. So GXO is a third party logistics provider. So they do order fulfillment for a number of different customers. And, one of them is Spanx. And Digit has been helping to do order fulfillment for Spanx orders at a facility in, Georgia for just over a year now. Just cross that boundary.

[00:02:28] Audrow Nash: So what does that look like? Does it mean that Digit is working an eight hour day, then taking a rest? Or is it working continuously during the day? Or how how does it how does it look for Digit to be deployed in this.

[00:02:41] Pras Velagapudi: A deployment looks a lot like Digit basically matching what the rest of the process can do. So in this particular facility, Digits working a full shift because the things upstream and downstream of Digit, the other stations are also working that same shift. So that's the limiting factor in a facility where there were two shifts, Digit could also work two shifts. So in this particular facility I think it's basically a single shift. I think it's roughly eight hours. Digit basically starts up in the morning and starts doing work where essentially it's a part of the process where orders have been picked from some, inventory upstream, and they arrive at a station on an Amr, an autonomous mobile robot that's carrying a bunch of different bins for the different orders. So Digit actually unloads those bins off of the Amr and puts them onto a conveyor so they can go to the next station, which is a packing station where they get packed to go out to customers. So it's one part of a process that interacts with automation on both sides. And I think that's really important is that we're building Digit not to replace all automation in the facility, but really to work with the other types of things that are there, like we have great automation for things like conveyance, and we have great automation for driving around a facility with an autonomous mobile robot. And then maybe not as great options when we're getting to a place where those two things interact and you're trying to get something from place to place, or you're trying to go from a shelf to a conveyor or a shelf to a mobile robot, or a mobile robot to a pallet that's going to go on a truck. Those are all places where you typically have to have higher cost, more customized automation that's built around what those other pieces are doing. So instead of doing that, we can have this single flexible platform that can go in and glue these other process flows together. And that's really the utility of a platform like Digit. We can come into a use case like that and without reconfiguring everything, without bolting a bunch of additional stuff to your armor, or completely reconfiguring your conveyor and adding a big gantry robot or something like that, you can just connect these pieces together and if you need to usage at somewhere else in the facility, the same platform, same Digit unit can go over there and basically with more or less mostly software reconfiguration, go do other tasks and that other part of the facility. So that's the real value of creating a human centric robot like Digit. It's to be able to be flexible and easy to deploy in all of these different places in a in a process flow that were traditionally very difficult and specific.

[00:05:26] Audrow Nash: Very interesting. Yeah. So, I mean, there's a lot of questions around why humanoids and I think that's a big answer to it in some sense, because what you're saying is that we have these systems that are highly optimized, but in the intersection or like in the boundary between them, it's often it's harder. And so like people would do it or you could have really complex, costly integrations that would be set up to make the systems map from one to the other. But the flexibility of the form factor makes it so you can kind of drop them in and they can do the task, and it's a software change. But I think it'd be worth talking more about this explicitly. Why humanoid for this? Like, what do you what do you say to people when they say, why not just a mobile base with an arm on it, like a single arm? Why do we need it to have legs? Why do we need it? To have the head? How it is like, what do you think is special about the human form factor? And why would this make good business sense?

[00:06:33] Pras Velagapudi: Yeah. So we often get asked why. Humanoid, right? I'm sure it's a very common question. And I think a lot of people are coming from the perspective of seeing companies that just arrive on the scene with a robot that looks exactly like a human.

[00:06:46] Audrow Nash: Yeah, you guys have been doing it for a while.

[00:06:47] Pras Velagapudi: Yeah, well, not only have we've been doing it for a while at Agility, you know, the company was founded. Yeah, back in 20, I think. About ten years ago. The company was founded a little over ten years ago. The design that we started with was was not very lacking at all. Cassie. Prototype, if you go back to those days, was essentially just a set of legs. And the legs look more birdlike than human, and it was designed to basically emulate a certain set of physics. So interestingly, with Agility, we've almost gone out of our way not to start out trying to aim at a human form factor. We've sort of evolved in that direction by trying to solve for the physics problems around interacting with things in human environments, and ending up with something that happens to analogously look relatively humanoid with some key differences. So you can see that, you know, our legs, for example, don't have the knees bending in the same direction as a human knee. Now, will that change in the future? Maybe because for our perspective, it's not about whether it looks like a human or not. We can we might not. It's about the physics of interacting with the world and how we do things like, the linkage transmissions to the actuators in our leg. That's what's driving the design. It just so happens that when you're dealing with things that are weights and payloads in human scale and around eyelets and things like that, that are again sized for humans, you start to up, you start to find that you're optimizing towards a human form factor. The legs end up being similar in size. In order to move human payloads, you need to be about human weight and human height. A lot of these things are converging, even though we start out saying like, look, make any arm of any length. And then it sort of ends up being kind of a similar length to a human arm to reach these types of payloads. You know why? Two arms? Well, a single arm can can contact a payload at a certain point. But the interesting thing about it there is that you have to exert all of your force around that one contact point with your payload. So if your pillow is really big, then what's going to happen is oftentimes you're going to grab it at the corner and then the rest of the payload is going to break off. Like, I don't know if you've ever tried to grab a really heavy cardboard box from just one corner. Not super great. And so you can either make a specialized gripper to try and exert torques and large forces across that, which is what a lot of special purpose systems do that are handling just a handful of boxes. Or you can add a second arm to hold another corner of the box. And now you're exerting really small torques comparatively, because you can do it at two points on the box.

[00:09:23] Audrow Nash: Now, could you do it with three?

[00:09:25] Pras Velagapudi: You could. In fact, some researchers have taken one of our research Digits, the, these three platforms that we produced several years ago. And they've put a third arm on it in a research application. But the more arms you put, the more expense you're adding. And so two is sort of our minimum. It's the least arms that can handle the range of payload sizes that we think is, yeah.

[00:09:47] Audrow Nash: Probably hit diminishing returns in terms of value.

[00:09:50] Pras Velagapudi: Right. And so you end up.

[00:09:50] Audrow Nash: With more arms. You go.

[00:09:52] Pras Velagapudi: Yeah you can make an octopus robot. But then that's a lot of complexity. You're trying to find the minimum complexity to solve the problem. And so similarly you know now we have now we have two arms because of the way that we're manipulating payloads. So why why legs. Why two legs in particular. You end up with a similar sort of thing. Right. You want to be able to dynamically balance loads in a really tight form factor. So you want to have the smallest footprint on the ground so you can navigate through these really small aisles, but be able to access a really wide workspace of the robot really high up, really low without expanding that footprint. So if you're doing that in a statically stable way, which is what a lot of the robot basis with arms on them are doing, then in order to do that, you have to make sure that all of the dynamic loads with your arms essentially keep you within the stability polygon. So the dynamic loads are like when you're moving a force back and forth, you have some weight and you're moving it back and forth. Yeah, dynamically.

[00:10:51] Audrow Nash: You want to get out of your support polygon.

[00:10:53] Pras Velagapudi: Moving. Yeah. You're moving your center of gravity. You're moving your center of pressure around in the polygon of where you make contact with the ground. And if you go past that then you start to tip, which is bad news, right? So if you want to do that well with the static base, you have two options, right? You either as your payload increases, you increase the weight of your base, which moves your center of gravity down and means that you move that center of pressure around less, less. Or you make your footprint really large, which makes your polygon big. And so you can move around more before you get to the edge of it. Right. But both of those are compromises that you have to make in terms of how much your robot weighs and how big it is physically, how wide it is now, with a dynamically stable base, you can have a very small footprint and dynamically change where your stability polygon is while you're manipulating stuff. So we can put our feet in very specific positions and push reaction forces through them to be able to maintain this really narrow footprint, even as we're moving through a pretty large workspace. So that's what we've found is a really valuable thing. It means that in a lot of different scenarios and a lot of different environments, whether it's narrow aisles or in situations in which there might be something on the floor that you're stepping around like a pallet, having the ability to really control that stability polygon and place where it's being, you know, where you're going to construct it and exert forces all the time to be able to basically control your whole body rather than just, react to the disturbances of the upper body, let's us do a lot of motion of payloads with a lot of dynamic load and a really compact footprint. And we think that's something that provides value. We think it's a reason that we're able to go into a lot of spaces where we can get really up close to things and we can access, you know, really large swaths of area. Now, there's still going to be a place for those static robots. And I would expect that we will see them coming on the market as prices decrease and things like that. But we were talking about what scales really well, and what can move from environment to environment. We're pretty confident that by starting with this base level of functionality, we won't have to adapt a whole lot as we get into more complex environments where there's more occlusions, where there's more narrow spaces or things that we have to fit in and around. Because again, yeah, we're human centric. So we're roughly going to be able to map to the same ground footprint that a human would need to move those same loads around.

[00:13:26] Audrow Nash: Yeah. Okay. Another one with this. And that's very interesting. All of that. And that makes a lot of sense. One other question that I hear often is why not have wheels on the feet? So you have legs. And so kind of like the Boston Dynamics, I forget what it's called, but has legs. And then it also has wheels at the feet. And then I think, I think you Unitree have some of these. Yeah. What do you think of this form factor.

[00:13:49] Pras Velagapudi: So we have seen it. So Boston Dynamics had the handle platform, which was kind of a balancing robot on two wheels. Super cool. Yeah. Very cool. And there's, I think, hexagon is trying to make a similar platform with two legs. We'll see how that goes. And then the Unitree, they have the go w which has the wheels at each edge at each foot. So combining wheels with legs, it's, it's an option that you can do. There's a couple of caveats to, to be aware of. And, you know, I won't say that, you know, Agility wheel will never explore having wheels at the, at the bottom. But there's some trade offs that you have to kind of consider. The first is if wheels are part of the dynamic stability, which is the case with handle or a Segway, or there's been a couple of companies at various points that have made, balancing, bots on top of balls. So, bossa nova at one point, for example, bossa nova robotics, and there's a couple of others. The challenge there is if you start to lose stability. So say you go over a slippery patch, right, and you lose contact and you start to lean the way that you have to correct a lean in a robot that's dynamically stable via, a rotary joint, like a wheel or a ball, is that you have to accelerate under the robot and get essentially catch up. The wheel inverted pendulum is where the inverted pendulum is. Exactly. And if you fail to do that, your only option is to accelerate more. And so, yeah, you end up in this sort of degraded state where the only thing you can do is to try and speed up the motion of the base. And the more that you're doing that, you're spinning up, obviously, the inertia there. And it's easy to get into a sort of unsafe, dangerous cycle where you're essentially accelerating the robot because you're destabilizing. And so the more you're falling over, the faster you have to go. And that is not a good loop when you're trying to get a safety case worked out.

[00:15:49] Audrow Nash: Yeah, that's very interesting.

[00:15:50] Pras Velagapudi: As someone who's face planted a Segway before, I can tell you that that's not a loop gets really fast sometimes. You were driving it. Yeah. Yeah. If you take a Segway sufficiently off road, you can get this to happen.

[00:16:02] Audrow Nash: To face planet. I had one where I was riding a segue, and, You reading a lot to keep it accelerating, and then it, like, kicks forward or something. That's it. Do you want to go faster? Yeah. Yes. So stabilize itself and keep keep it, keep it. It's, center of mass under its support polygon, I suppose.

[00:16:20] Pras Velagapudi: Right. And so if you lean really far forward, right, it loses the ability to go fast enough to keep up with that. And that's when it starts to get a little bit hairy.

[00:16:30] Audrow Nash: So wheels basically make it so it's hard to gear. It's harder to can think about safety for this because you have to outrun. It's falling over if that occurs.

[00:16:39] Pras Velagapudi: That's specifically in the case of when they're a part of the dynamic stability. Now you have other robot platforms. And a good example of this is the, the Unitree, go where they have the wheels just on every foot of a quadruped. So, so they're, you're really just if you're just using wheels to create, fixed velocity on your otherwise dynamically stable base. Right. They're using the legs to do all the dynamic stability there. And that's a very different thing. That would be the same as if you, for example, made like roller skates on each foot of a, of a biped because they're you're still using the fetus feet. It's just that you can also induce a velocity on them. So that doesn't necessarily have the same constraints because it can degrade by breaking all of the wheels together. You're basically just turning it into a foot. But it's a lot of extra joints and a lot of extra power and mass that you're putting at your feet, which ideally, the less mass that you can put in your foot, the faster you can move the foot to do the stepping. So in some sense, if you make a really heavy roller skate or like a skateboard, that you're riding all the time now, if you do need to move your foot to do, you know, stepping things, you've put that much more mass on it. So it's a bit of a trade off that you usually don't want to take on until everything else is sort of solved for right now.

[00:17:57] Audrow Nash: So in the environments that you're in, that Agility and Digit, that Digit is and are the floors rough or anything like this? Is it like I'm wondering about how that trade off would be, because I get that. So if you, if you treat if you have wheels that act as roller skates and you are just kind of using them to move faster, but you can turn them off to have them brakes, so then they're just regular feet. I get that they are heavier to pick up and thus you lose more energy when you step. But, would you lose less energy if you were just traversing on a flat floor with thumb, or how would you think of that kind of thing? Yeah.

[00:18:37] Pras Velagapudi: If you're if you're just comparing wheels to legs, when you're moving around in a flat environment, wheels will be more efficient. They, they are for humans as well. That's why we have things like bicycles and and so forth ourselves. Right?

[00:18:50] Audrow Nash: Hoverboard and whatever else. Yeah, it.

[00:18:51] Pras Velagapudi: Shouldn't necessarily be that much of a difference in, short distances. And when there's a lot of things like turning and reorienting involved there, if you're doing an efficient walking gait and again, you can do an inefficient walking gait. This is something that you have to actually design for. You want low mass in your legs. You want to have a good control algorithm that's not making huge static forces in your toes and things like that. But once you get all of that right, it's not too much worse to be doing walking in these small areas like turning around in small, you know, work cells and in between hotels and things like that. Over a longer distances, a wheel is a little bit better now where it's more interesting is actually not in the hey, I need to step versus, drive and I am using a little bit more power. It's I unexpectedly need to change my stability polygon because I, you know, maybe misstep or I clipped some object, right. Like, I maybe I'm, I'm, walking along or you're driving along, and I encountered the edge of something like an anti fatigue mat, or someone has left something on the floor. Right. Yeah. With Digit right now will very quickly be able to do what I call step recovery, right where we essentially sense that this is the disturbance and start moving our foot really quickly to a new point, you know, go over or around the obstacle to a new point where we can reestablish our stability. Polygon. If you create a lot of mass on the end of your foot, it actually changes.

[00:20:22] Audrow Nash: How slows it down, move it right.

[00:20:24] Pras Velagapudi: Do you have a lot of inertia to move? So now you need bigger action to do that response, a lot more power to do that response. And that makes it harder to do so. Having faster feet really lets you be very responsive. And that's the bigger thing that you're losing when you put a lot of mass in the feet is actually that things like step recovery, you're now going to use a lot more power and you're going to be sluggish to respond. In the same way that if you're a human and you're wearing like really bulky roller skates, it's definitely going to be easier to trip you up than a person who's just going around and, you know, like cross trainers or something.

[00:20:58] Audrow Nash: Yeah, for sure makes sense. Yeah. Because if I think of a lot of times like if I trip, I'll like run out of it, almost this kind of thing. And that's because of that fast step that you can do to recover. And that's what the robot needs to do. And if you have super heavy feet, that doesn't work very good. What do you think about the idea of, like, a humanoid torso on a centaur, kind of or on a, on a quadruped base? Yeah, it's kind of thing because it seems like you could have a lot of. You could have a lot of the benefits of legs and even wheels. But then you don't have things like these, the, the tripping issue as much because it would be better if you had four legs if you were to trip. But I guess it's more parts and it probably can't turn around as easily in narrow spaces. But how do you think about that kind of thing? Because a lot of people have thought have mentioned that like, a centaur shape seems like a very optimal shape.

[00:21:51] Pras Velagapudi: Yeah. I think you're more or less nailed it with the footprint really being the thing that becomes quite inconvenient. So if you start to build a four legged platform with a with a torso on it, you start to get into the same degradation of, you're starting to really take up a lot of, footprint on the ground in order to do that. And that really impedes you being able to get into some of the narrow passages that humans are expected to operate in. And therefore, robots like Digit are expected to operate in. When you're doing things like going between a conveyor and a shelf down in Iowa or something like that, you wouldn't really be able to turn around a centaur robot, as effectively in those spaces. And it's not too different once you do all of that of versus what you might need to do, have a statically stable wheeled base in kind of that same form factor, you're you're also taking up, you know, similar expansion of the ground footprint. So the sense of a robot is sort of an interesting point in the space. It's it's maybe not clearly so much better than just doing that with a statically stable base that's that large, if you're going to do it anyway. Folks will explore it. But really, that isn't necessarily a great trade off if you're trying to get into, getting up close to different types of machinery and applications. That for the same reason that it's not super, convenient to, you know, take a small horse through an office building. Right? Yeah.

[00:23:15] Audrow Nash: No, but it'd be great to see. That would be funny. Okay. Hell, yeah. So tell me more about Digit and the applications that you guys are using it for. So it's a lot of code manipulation and moving objects. That the last time I saw I thought Digits hands were flippers. Like, they, they only had one degree of freedom. Maybe they have a wrist actuation, too. So two degrees of freedom on the hand. But you have their Digits. Design is modular so that you can swap out parts. But tell me about the tasks that Digits doing and I guess the end effectors and how it's equipped to do those.

[00:23:52] Pras Velagapudi: Yeah. So right now Digits primary use cases are all material handling. It's where we wanted to enter the market. So Digit right now has for user field replaceable units that are essentially the limbs of its body. Its arms and legs are the field replaceable units. So we can swap out an arm. And usually we just swap it out with another arm. It's, you know, service operation. But when we want to change functionality of an ARM, we can swap out an arm that has a different end effector on it, for example. And so the, the, the end effectors that you're talking about, that sort of flipper style of end effector, we're actually from an older generation of Digit that was, moving around types of cardboard boxes and certain types of very, friction based, material handling contact points. So it kind of clamped down on the sides of that, what we've switched to and have been using in most of our deployments now is actually what looks like almost like a pincer. So it does have it has essentially four fingers, two pairs of fingers and thumbs that grab something like a tote, at sort of a front and rear, pinch foods. And so those work really well for things like bins and totes and other places, other types of containers that have an overhand grasp position. And so we've been doing a lot of that type of material handling. Those grippers have been the ones that we pretty much most concern most things. And those are the ones that you can see in our deployments to conferences like, Pro at this year where we were showing off our new Z4 production units with, the extended battery packs and a new version of that, a poseable grasping gripper with the four fingers. So that's what we've been deploying on, on those robots, primarily as we're going forward. What we're planning to be designing into the next version of Digit is actually just a tool flange. So Digit can just change out its own end effector super as it sees fit. I mean, that's where the industry has been for industrial manipulators for some time. Yeah. So it's not really a surprise that it makes sense for us to pursue that basic awesome subdivide our our field replaceable unit into an arm and a hand. Essentially, although awesome hands might not be a hand.

[00:26:07] Audrow Nash: It might be a.

[00:26:08] Pras Velagapudi: Claw or a hook. Or maybe.

[00:26:10] Audrow Nash: A hand. Could be anything. Yeah, could be a saw. Could be a solder. They could be anything. Yeah. Yeah. Super super cool. Yeah. I love the idea of, like, tool changing robotic parts. Well, the robot's doing that. Like, I don't need that anymore. It switches its hands, grabs a more optimal hand. So cool. We're in the future. It's unbelievable.

[00:26:27] Pras Velagapudi: It's really funny because a lot of, a lot of times you see people really thinking about, you know, these five fingered hands and they're trying to jam just like, you know, 20, 30.

[00:26:35] Audrow Nash: Did incredibly complex that.

[00:26:37] Pras Velagapudi: You know, exactly replicates what our hands do. But if you really think about it, there's there's two kind of things there. One is if you actually look at how people grasp things, it's not like they necessarily use all of their degrees of freedom for everything. And you have a number of folks who are highly capable, who don't have all of those degrees of freedom for various, disabilities, medical conditions or things like that. Right? Whatever. Good. But they're highly capable, and even people using various types of prosthetics can be as dexterous as any robot on the market today. And so it's clearly not that you have to exactly be replicating a human form factor to get to dexterity. It's there's some notion of dexterity, which is helped by having some part of those degrees of freedom, but it's sort of an open question of what's the important bit there. And certainly now that we have AI based models for training it, we can explore that space in a lot more, effective way than we could before by learning about how to grasp things with various types of gripper configurations. And then.

[00:27:35] Audrow Nash: The oh, you mean you mean making a simple gripper go a longer way through, I don't know, reinforcement learning or something like that, where it goes and iterates and tries everything and you see what you can do with it.

[00:27:44] Pras Velagapudi: Right? If you look at a lot of the work that's, coming out of groups that are that kind of started with the, aloha work. But now, you know, you see companies like physical intelligence, and Dana really showing off that even with simple two fingered grippers, they can do a lot really complicated folding, manipulation and things like that. Really, I think there's a lot of, a lot of, low hanging fruit available in being able to unlock dexterous manipulation with what we can do with relatively simple grippers. Now, I think that the same techniques might also unlock a lot of utility out of really complicated hands. But really complicated, multi-step hands means a lot of failure points. It means that when you move to high, low.

[00:28:27] Audrow Nash: Very expensive, those degrees of.

[00:28:28] Pras Velagapudi: Freedom are now expensive because they.

[00:28:30] Audrow Nash: Have to.

[00:28:30] Pras Velagapudi: Deal with dealing with the large forces of a big payload or something like that. Totally. So I think there's a there's this notion of dexterity, and we all agree that more dexterity is more capability, but does more degrees of freedom equal? More dexterity is maybe an open question versus maybe you just need like a gummy hand pad or like a big clamp thing, you know, that's that's sort of open, right?

[00:28:52] Audrow Nash: Yeah. Yeah. I think the thing that I always think of when I see those, like, really ornate hands with all the degrees of freedom, it's like, I don't know, I, I heard it some friend or something was like, it just keeps nerd sniping people where they, like, want to go build the perfect hand replica. And it's like the utility of it's not that good, but it's a very interesting problem for this kind of thing. Is fun to work on. Yeah, but, it's a beautiful engineering problem.

[00:29:20] Pras Velagapudi: Right? But then when I go take their hand and try and pick up a 50 pound tote and break all of the fingers off of it, and then it's like, yeah.

[00:29:27] Audrow Nash: And then it's so expensive and time consuming to repair and repairs take a while. So yeah, we'll.

[00:29:31] Pras Velagapudi: See where that is. But there's a second piece to it, right? Is that there's plenty of things that we do in our own environment where we really could do it a lot more effectively if we had tool changing, like if I'm using a drill right, the entire handle assembly and trigger assembly and speed setting and all of the features of that drill that are just there because a human can't plug into it and have a direct IO connection that says drill, please turn it this speed with this torque.

[00:29:58] Audrow Nash: So freaking cool.

[00:29:59] Pras Velagapudi: Yeah, it's just that's a sort of local problem to us, right? Like a robot doesn't have to have those problems and so we can just engineer our way around them. Right.

[00:30:11] Audrow Nash: Super, super cool. Yeah, it's really interesting because it, I guess it's cool to me to see robotics going from very specialized towards the application to more general form factors now, like humanoids. And then it's super cool to see the ability for humanoids or whatever form factor to do tool changes, to better suit itself to whatever the task is. It's doing super, super cool.

[00:30:36] Pras Velagapudi: Yeah. I feel like we're seeing the convergence of all of these different seeds of technology that have been advancing in the background for years. So things like energy storage getting better, things like actuators getting more power dense and more, mass efficient. And of course, on the compute side, just more powerful compute and better AI modeling and techniques to be able to control complex systems and interpret complex perception data. All of these things have been sort of raising their water once and for all, crossing over to create this new generation of much more complicated physical systems. And so it's really cool because essentially the water line is just rising, and we know that we can do these things now. And it's just created this new level of capability that that I think is really going to change how we approach automating stuff in the future. We just can do this now.

[00:31:29] Audrow Nash: Yeah. I think the thing that's interesting to me is it's like before it felt like, oh, that's not feasible. This won't work now. It's like, it's not efficient kind of thing, which is an interesting thing.

[00:31:43] Pras Velagapudi: Yeah, it's really funny because having gone through that, you know, Agility sort of went through that path where when Agility first started, you were convincing people that a humanoid could even work and you'd be like, come to our facility like, look, we'll show you one that works. And now it's a lot of people complaining that they work fine. They totally believe that. They're like, coming to our facility. We 100% believe you can solve this problem. And of course, they'll give us some incredibly complicated, very difficult problems. And then they're like, why aren't you doing it faster? Like go faster, be cheaper. And then you're sort of like guys like five years ago, you wouldn't have even believed that you could do this. And now you're complaining that it's not good enough and you should be doing it.

[00:32:20] Audrow Nash: That you know. Yeah, it's an interesting thing. Let's see. So, well, tell me, you alluded to it a little bit. So you have your version four robot. And so that's the one that you're doing now and that's that one that has that interesting gripper that you mentioned. And you mentioned that you have a version five coming, didn't you, before, and you mentioned a little bit, tell me a bit about that and tell me what's changing.

[00:32:47] Pras Velagapudi: Yeah. So with the version five, the big changes, version four is designed to work out in industry and be compliant with regulatory standards for working in a facility with an external safety system with some form of a work cell around it. So we need to separate the robot from humans. Now, the robot has an onboard safety system, but it doesn't necessarily have all of the onboard capabilities to detect humans and to be able to dynamically change what it's doing in response to humans getting close to it. So we need to either put up walls around it, or somehow segment where the robot works from where humans are. And if a human enters the space, we basically kind of shut down the robot before we let the human enter the space. Our big design change with our V5 platform, which we're working on right now, to come out in about 18 months or less, is to be able to integrate all of that safety onboard the robot and make that robot safe and compliant to operate cooperatively around humans, which is to say that the robot and humans can move through the same spaces together without specific guarding or protection around them. All of that's included onboard the robot, so they can't necessarily work on the same thing at the same time. That type of collaborative safety night, because that requires a little bit more than than what exists right now in the safety industry. It requires some fairly advanced technology there that doesn't quite exist. But pushing what we can do now to its boundaries, we can do cooperative safety, which is that the robot can detect humans in its vicinity and limit its motion with respect that using onboard safety systems. And so that's really pushing us to do this, reinvention of the Digit technology stack to incorporate safety at every level, from the physics of the robot itself all the way through the drive systems, the drive electronics, and the low and high level software to build this new platform. So it's pretty exciting. And we've been doing a lot of work, not just from the robot design perspective, where we've been developing out a lot of new technology for, AI based safety and on board dynamics stability, but also on the safety side itself, where we're also developing out new standards that can use these types of models as part of a safety strategy that we can take to, a nationally recognized testing laboratory to basically, guarantee and certify that our robots are in compliance with industry standards. So that's been a huge effort as well. And we're pretty excited about that. And we we just recently, basically, created a new, ISO standard that's in development. So we're now working on a specific ISO standard towards safety for industrial robots that use dynamic stability, which is robots like Digit, where the dynamic stability has to be part of the safety strategy.

[00:35:54] Audrow Nash: Okay. Tell me about that.

[00:35:56] Pras Velagapudi: So in previous standards, things like NCR 1508 which deals with autonomous mobile robots, a lot of the standards, basically have gaps in them when it comes to how they're applied to a robot that's dynamically balancing. They kind of make the assumption that you more or less probably have a statically stable robot. It's either bolted to something or it's on wheels or something like that.

[00:36:20] Audrow Nash: Because that was true in the.

[00:36:21] Pras Velagapudi: Past, right? It was true for industrial arms. It was true for most collaborative arms and and for amrs autonomous mobile robots, which are primarily wheeled bases that have, you know, shelves and things mounted on them. So that means that a lot of the standards essentially say, oh, well, in order to guarantee safety, like, here's how you're going to power the entire robot in this in this type of scenario.

[00:36:41] Audrow Nash: Safely, right? Yes.

[00:36:42] Pras Velagapudi: Which works if locking up your wheels is a thing that makes you more safe. It doesn't work. If turning off all of your motors is a thing that makes you so that you can no longer keep yourself upright, which is the case in any dynamically stable robot. So that's the case in Digits, certainly with legs. But it would have been the case, for example, in the Boston Dynamics handle robot. Or if you were trying to make a Segway that could work, you know, with an industrial arm on it or something like that. That's also a dynamically stable robot and would have some of the similar gaps in specification in the safety standards. So this new effort, which we're doing in collaboration with other folks in Boston Dynamics, is also participating, with, with our tool, it's a collaborative effort with a bunch of different, providers that are working on these types of platforms to basically to find out how do we do this dynamically stable safety. What do we need to, specify? What do we need to, describe so that we can ensure that a dynamically stable robot that's doing certain things at certain levels of reliability could be considered safe enough to operate in an industrial environment. And it really is kind of like clarifying some of those details to get this to match with broader, you know, industrial standards for safety of machinery as a whole. So all of these standards are kind of building up. So there's, different levels of standards like type A through type C, and the idea is that you have broader standards that essentially say like make this robot safe and then more specific ones which are around. And here's how you make a wheeled robot safe, and here's how you make a collaborative robot arm safe. And now here's how you make a dynamically stable robot safe.

[00:38:27] Audrow Nash: So how do you make a dynamically stable robot safe.

[00:38:31] Pras Velagapudi: Yeah. It comes down to a few things. And there's a little bit of secret sauce in there that we're, that we're building out. So I can't talk about all of the detail. But what we think is kind of primarily important is that you need to understand where people are in your environments, and then you need to be able to be maintaining, essentially, a safe operating envelope for your own robot that you can guarantee in the events of faults or failures. That's within the space of where people are. So you have to essentially say, oh, how do I keep my robot such that it's bounded in how it's going to move or exert forces, or even trip and fall and things like that? So it'll always end up within this, this bounding radius or this bounding sphere or whatever it is. And I can make.

[00:39:20] Audrow Nash: Sure that there's no person.

[00:39:21] Pras Velagapudi: That people, as they're approaching, I can shrink my radius down by doing things like slowing down my motion or getting closer to the ground. And so it's this whole dynamic process of how do I enforce that as a person gets closer to the robot, the robot is reducing its hazard by the things that it can control about it's balancing and its posture.

[00:39:43] Audrow Nash: That sounds really cool because so I'm imagining that a robot is moving through a space and it's estimating where there is risk for it falling and doing damage. And what it's doing is it's making sure people aren't in that zone for this kind of thing. And there's probably some risk negotiation there where it's a tight space and I have to. So I'll go real slow or something like that. Is it it's similar to that kind of thing. Yeah.

[00:40:09] Pras Velagapudi: It's this idea of basically trying to figure out where your various levels of hazard can be and then, yeah, what can you do to, you know, if you got to go faster and increase that radius or you can, you know, do things like go slower or use lower forces and shrink that radius. And now if you twist that lever, right, if I can combine that with detecting where people are, and I think detecting where people are is kind of an important piece here. Yeah. Because you're on a platform whose job. Right. The thing that you're getting paid for to do is exert large forces on objects. Right. I need to pick up a 50 pound tote. That's a huge force. That same force, if I apply it.

[00:40:49] Audrow Nash: Did you pick up a 50 pound tote?

[00:40:50] Pras Velagapudi: So the newer versions of Digit can pick up, up to because that's heavy. Yeah. That's impressive. 23-20 five kilograms. That's what we're targeting for our V5 platform. The existing robot can within a higher workspace do 16kg. And it's higher as you go a little bit closer to the robot.

[00:41:07] Audrow Nash: 30 pounds.

[00:41:08] Pras Velagapudi: Yeah. Like 33 pounds or 35. Yeah. And it's doing that. It's doing that, you know, all day every day at facilities. It's actually handling those loads.

[00:41:15] Audrow Nash: It's awesome. So yeah we 50 pounds is all we.

[00:41:18] Pras Velagapudi: Have Digit you know. Oh yeah doing, doing little dumbbell curls with sandbags and things like the.

[00:41:22] Audrow Nash: Barbell and whatever. Hilarious. I would love to see those videos. That's great.

[00:41:27] Pras Velagapudi: Yeah. Okay. So basically once you're doing that right, you can't you can't limit yourself just by saying, oh, I'm really close to something that could be a person or a shelf or something. I need to turn down all of my, forces.

[00:41:41] Audrow Nash: That'd be useless.

[00:41:41] Pras Velagapudi: And you'd never be able to carry anything exact for sure. So that's why understanding where people are is such a big part of having a functional robot that can also be safe.

[00:41:53] Audrow Nash: So that sounds like a lot of perception is required. I'm imagining almost like like I'm thinking of Tesla and their cars and how they see everything around. And so I'm imagining that you need something similar. So the robot has no blind spots, so that you can go and look at everything and assess where things are, or you need a good map of the environment, but then always a person could sneak up behind you if, say, you only have forward facing eyes and that would make your estimate bad. So do you basically have like super aware perception, like, is that part of the standard or tell me about the perception side of this. Yeah.

[00:42:33] Pras Velagapudi: So this there's two pieces here to to making a safe robot. The first is the standard where we're essentially writing the rules. So in that part of things we're not really defining like should you use cameras or should you use later lidar or things like that. You're really defining like what would you need to prove in or and what would you.

[00:42:50] Audrow Nash: Standard as modular? In a sense, because of that, you're making there's no assumption on how you get this data about the environment. You just said once you have it, this is the case for safety.

[00:43:00] Pras Velagapudi: Yeah, it's it's defining. How do you go about, how do you go about guaranteeing that that are dynamically stable? Robot is safe like what's what are the rules. What are the what are the checks that you need to kind of meet. Right. But not the technology. And then on the other side, you know, at Agility, we're obviously developing our platform here to try and be able to meet those rules. Right. And that's where we're deciding things like, okay, how many cameras or other types of sensors might we need? Where would we place them. You know, you can you can, for example, choose to, have fewer sensors but then need to maintain more state about your environment. So essentially you could say, well, as long as I remember that I passed by this area and a person wasn't there, I could assume that someone couldn't have snuck in, because I can see around all of the places that they could come in from, even if I can't see that specific spot. Right. So there is some amount of that. But then you, you then if you lose that line of sight or you get that, or.

[00:43:58] Audrow Nash: They were hiding in a shelf or something.

[00:44:00] Pras Velagapudi: Yeah. So then there's edge cases you have to deal with where you have to maintain that state. And that might be better served by just adding an additional sensor there or something. Probably in some cases.

[00:44:09] Audrow Nash: Yeah.

[00:44:09] Pras Velagapudi: Clears the space for you. But you have to make those trade offs right. And so we do plan on having a bunch of sensors. That's I think an important part of this. But there is also some amount of understanding exactly how the environment evolves from a safety perspective and clearing certain areas or reasoning about where could a person be realistically, given that this is what I've been doing? Like if I was just in that space and I could see all of the areas around it that a person could have moved from, so a person couldn't have, you know, teleported in there, right? They would have to have already been in there, in which case I would have seen them before or after something like that. Right. You can have that type of reasoning now. Yeah. For sure. It'll be a combination of, of some amount of understanding that from the safety side. Right. And also just having really good sensing and a variety of sensing that can reliably detect humans under a variety of circumstances.

[00:45:05] Audrow Nash: Yeah. And one thing that's really interesting to me with this, from how you described it, I think you said it like on agent or onboard safety for this kind of thing as opposed to with cameras everywhere. And that gives the robot the perfect, observer that it gives the world perfect observability for the robot. And maybe that would work. But it's it's less flexible in its ability to deploy because you need power and data, ability to stream data if you have cameras everywhere.

[00:45:32] Pras Velagapudi: So you can actually pair even the V4 version of Digit, right? So the v4 version of Digits designed for an external safety system, you can pair it if you have one with a safety system that's using external cameras or external, laser or in various forms, whether it's 3D lidar or things like that. And people do make systems for doing external safety where they are detecting people from a third party, not a third party, but from an external, perspective. You put cameras all through your facility or some area, right? And then you, you essentially detect where people are and where robots are. And you can use that to then write a rule that says, hey, if something's getting close to a V4 robot, shut it down, right. So you could do that. Now, it's just that those systems require you to build out that infrastructure, which a lot of our customers are coming to us because they they don't want to have that infrastructure in any specific place because they want Digit to kind of come into an area and be ready to go and maybe move it to someplace else later. And they don't want to necessarily put a camera system here and a camera system there.

[00:46:37] Audrow Nash: And you see you have a 30,000 square foot warehouse or something or three, like some massive, 1,000,000 square foot warehouse. It's like, that's going to be pretty costly to put cameras everywhere.

[00:46:47] Pras Velagapudi: Yeah, you might be able to do it in a certain work cell area so that for that use case it might make sense. And when you can do it, if you can afford the cost because the systems are not cheap, you can certainly do it. And it might be even better than having onboard safety. But it's really about, again, the infrastructure versus the flexibility.

[00:47:06] Audrow Nash: Yes, definitely. Yeah. And you guys are heavily betting on the flexibility, which I think is a good bet in all of this.

[00:47:13] Pras Velagapudi: It's it's a part of the space. There's always going to be some part of these industries that needs that flexibility, whether it's a type of industry where the amount of flow changes really seasonally. You know, one of the reasons that grows so interested in us is as a third party logistics provider, they get new customers all the time. And so a single facility might be changing what fulfillment it's doing, or might add or remove a customer throughout the course of a year. And so you really don't want to be in a situation where you paid for a whole bunch of automation that's at a single rate or a single size, and now I add a customer and it's now the wrong size, and I have to rip it up or add a really big expansion or something. So that flexibility for them is really valuable. And there's other cases in which you have the use of that flexibility is because you might have a really big facility. We're using different parts of it at different times. Right. You might do something in the morning and something else in the afternoon. Or there's things like cruise shifts where a single person rotates between different stations. So that's really common in certain types of manufacturing. So I have another customer, Scheffler, who does automotive parts, where that's a really big part of where we can bring value to them is that they have, you know, parts, machining equipment. So things like washers and dryers and heaters and things like that, heat treatment systems that basically batch up work. So you load them up and then you leave, right. And they go, you know, clean parts for 90 minutes or something, and you don't need to be there for those 90 minutes, and it'd be a.

[00:48:44] Audrow Nash: Waste to.

[00:48:45] Pras Velagapudi: Something else. Robot arm to be sitting there for 90 minutes, after 90 minutes, then a bunch of bins come out and you need to take them and load them into the next thing. Right. And so they have people that often do these functions right now. So it's really a very useful platform for being able to go between these different stations that need that type of service, but not be there all the time.

[00:49:08] Audrow Nash: Yeah. Where you can go have them do something else.

[00:49:11] Pras Velagapudi: Right or man multiple stations or something like that. Yeah.

[00:49:16] Audrow Nash: So one thing that I'm thinking with your so it is a new ISO standard, a new international standards organization standard. Why make a standard I so the way I'm thinking, which I'm happy you guys are making a standard because I think it'll improve everything for everyone, but, why? Like I'm imagining it could be a competitive advantage. I don't know what the other humanoid companies are doing or dynamic robot companies. But. And maybe having a third party validate things. I. Can you tell me some of the strategic thinking behind it? Because I'm wondering if you didn't make a standard and just had an internal secret sauce for these safety things. Maybe it would be a significant competitive advantage. But how do you think about this?

[00:50:06] Pras Velagapudi: Yeah. So this is standard. I mean, to to be clear. Right. We still need the secret sauce about how us to do it. And anyone else that's trying to do it is is going to achieve the standard. The standard itself is laying out the rules, right? The rules for what you need to do. So what they're.

[00:50:21] Audrow Nash: Probably well thought out. And that saves a lot of early work and back and forth with customers and things like this, which I'm probably, I'm sure you did in creating the standard.

[00:50:30] Pras Velagapudi: Right. But one of the reasons that we're pursuing it is because in order to deploy robots outside of a work cell, we already need to meet standards, right? So within the US, you know, OSHA, while it doesn't specifically, you know, name a certain standard, is basically putting out the expectation that any automation deployed within the US is going to meet a set of prevailing standards. And so you already have ISO standards for things like the safety of machinery, the placement of stocks. And and you're already required to be in compliance with them. They just don't have good specifications for what it takes for our robot to be in compliance with them. Yeah. Outside of a work. So so it's really that in order to be in production in the US and especially in Europe, where it's a lot more important, it's basically a requirement for you to be able to certify your robot at least as partially completed machinery, to go into a production environment. Well, you really need to have the document that specifies what you need to do to meet that, right? Because otherwise you don't really have a path forward to to say, hey, I'm I'm meeting this this requirement. I'm, I'm complying with, industry standards for best practices here because there, there aren't really completely defined industry standards. So for us, it's really important that we get these standards out there and and written and specified so that with our customers together, we can go out and ensure that we're in compliance to meet the legal obligations that we have, when operating in the U.S or abroad.

[00:52:08] Audrow Nash: How do you think it works with other humanoid companies? Well you say like figure with their I think BMW partnership for Tesla with their one or the Tesla or one X or.

[00:52:19] Pras Velagapudi: Yeah. So it right now the way that pretty much all humanoids can be deployed is is kind of in two ways. One is within the US you can deploy them you know not it directly into production applications. You can deploy them into basically research environments or other types of new innovation spaces. And those don't have the the same standards requirements. Obviously, you still want to make the robots safe, but you're not being specifically directed to like an ISO standard for, for an industrial standard, if you're deploying it in essentially like a research lab or an innovation part of a of a facility. And so in a lot of cases, and certainly even in our earlier POCs, we'd often be deployed into an innovation center or into a special facility that has some portion of the facility essentially running non-production workflows. And the difference there is that when you're deploying in a research environment, the requirements are a little bit different because your exposure is a little bit different in terms of the people that you expect to be around the robot, and how the robot's being operated. When we're talking about why it's so important to meet industrial safety standards, part of it is because once a robot's deployed in a facility working full shift work, the types of people that are around it and their level of training with respect to the robot is different, and the amount of attention that they're paying to the robot is different. These are people who are now they're not there just to work with the robot. It's not like an innovation center where the whole point is to look at the robot and see how it's doing. They're trying to do their own work, right? Their job is not to follow the robot around. It might be to walk past the robot and do something else. Right. And so they're not necessarily paying the same level of attention, certainly over 8 or 16 hours. And so for sure, that's why it's so much more important that you're ensuring that the robot performed a certain level of reliability and how it's it's doing the safety. So there's there's this one aspect which is that, you know, you can deploy into different environments that are not the industrial production environments. And a lot of people start out there. The second piece is that you can deploy humanoid robots, and we're doing this right now in conjunction with the work. So if you have another system that's around the robot, that's really what's keeping humans safe. And maybe that is plexi glass panels or, a light curtain.

[00:54:37] Audrow Nash: Or something that bounds it.

[00:54:38] Pras Velagapudi: Then you can do it right. You can put that around the robot and then achieve, some, compliance with safety standards, at least in North America right now. And that's why you see that that's what a lot of people are doing. They'll deploy into parts of facilities, or they'll deploy in use cases where the robots really, not moving between stations or it's moving around within some restricted zone. And that's how we operate right now at GCC. So we have a restricted area that the robot operates within. And it has a link to an external safety system that helps guarantee that we can shut down the robot from outside of that work cell if we need to go into it.

[00:55:16] Audrow Nash: Cool. Do you do you know if any of the other companies making humanoids have their robots in production? And if they do, they're probably in work cells, I would assume. But yeah. What do you what do you think is the current state of humanoid robot companies other than Agility?

[00:55:34] Pras Velagapudi: I think that it basically because of the way that the standards are defined, everybody has to at best be in production in a work cell, or if they're not in the works, they'll be in a non-production environment. And in general, that's just everybody has to follow the same rules there, at least within North America. And Europe. Yeah, right. I think we often see, you know, one off, demonstrations of technology. And certainly we can teach them to. Right. Like we did a recent demo for, for Google, where we work together using their new Gemini models to go pick items off a grocery shelf. Right. And that was in our innovation space. And the robot, you know, grabbed a little basket. And I think I think we told it to get the ingredients for spaghetti or something. And so it's picking tomato sauce and pasta off the shelf. Right. The robot could do that. Right. But it can't do that in production. Right. That's not a production capability. And that's in part because that's not something that it could do, without a work cell around it. Right. And that's the same for us as everybody else. Anyone who's doing this legally has to be following the same sort of practices.

[00:56:43] Audrow Nash: Make sense. So you needed to do this to enable a lot of high potential things for you guys to do this, to create the standard, because there was a gap in the existing standards, because these robots are so new. And with this kind of thing, with this additional standard, once you prove that once you write the standard, then prove that you need it, then you are able to have your robot expand it to a much wider, number of applications, probably, or a much larger application where you are working outside of a work cell for all of this.

[00:57:18] Pras Velagapudi: Yeah. And that's exactly why we went after it, right now. Because what we're really seeing is that as we're moving into the space, there's a lot of capability that that we can show that others are showing of the platforms. You know, we could pick groceries off of a shelf. You know, we can we can walk around outside, you know, Digit can walk around outside. Yeah, totally. We don't have the safety rules about how one does that for a lot of these environments. So we're really seeing that safety is a key gating item for getting dynamically stable robots out into the world. It's actually probably going to be, lagging behind these capability of these to do things in space. So we'll we'll be able to pick groceries off a shelf. We can pick groceries off a shelf, but we have to come back with, well, what are the safety, safety cases for doing that in a retail environment? You know, what are the safety cases for doing that in a construction environment? In a home eventually. Right. Those have to be worked out. And homes are particular challenging. They're sort of hardest case because not only do you have all of the lack of structure of an alarm about your home, but mine's not the neatest. Right. Chaos. Right, exactly. But, you know, two to exactly your point you were mentioning earlier, you've got a small child and you've got an animal in your house, and yeah, in both of those, as a small children and animals. Pets, right, are things that you have to be safe around. If you were a robot and you can't expect them to either be trained or avoid a hazard, right? You can't be like, no, we trained all of the puppies to never come close to the robot like, you can't, I don't know, I, I have to make my robot never do anything that could cause a hazard to your puppy, right? That's the way. Oh, yeah. Right. And that's a pretty high bar. That's way harder than in a in an industrial facility where.

[00:59:06] Audrow Nash: Very much pretty.

[00:59:07] Pras Velagapudi: Much adults, they are probably receiving some sort of training. There might be PPE restrictions on certain areas anyway. And those are reasonable. What's called administrative controls that you can put in your robot still has to have a certain level of safety, but you can do that in conjunction with some of these administrative controls. And that really simplifies things quite a bit. Right? Yeah. In a home like you have to be taking care of you, everything has to be on your side. You can't induce a hazard that could, that could harm any, any, any of these individuals. An adult, a child, a pet. And you have to assume that they are not necessarily doing anything to make your life easier.

[00:59:48] Audrow Nash: Yeah. Very much. We also have chickens and those would be chaos near the robot. And it's funny because they go to you when you go outside. So they would go right to the robot probably.

[00:59:58] Pras Velagapudi: Yeah. Exactly. Just running around its feet you know.

[01:00:00] Audrow Nash: Hey hilarious. Okay. That's a very interesting thing. So do you think that how do you imagine these safety standards? So is this ISO standard, are you working on it or is it done or.

[01:00:18] Pras Velagapudi: So we just got what's the status of it. Yeah. We just got approved to create the the working group for it. So now there is a working group, a team, an international team of ourselves and other, robot companies and other industrial automation companies that are participating to to define it and build it and ratify it. So that's that's where we are right now is we've we've got the group together. We've got approval to go after this. Everyone's recognized that this is an area that needs a new standard. And so now we're building it out together with, other folks in the industry.

[01:00:54] Audrow Nash: What does the timeline look like for this? I guess the whole thing would be interesting to getting the working. I guess it's the that's logistics and you just have to be approved. What what's the timeline look like for this?

[01:01:08] Pras Velagapudi: Yeah. So it does take a it does take a while. We do have the benefit of, one of the key folks on our team is Melanie Wise, who largely did this work before for the AR 1508 standard, which is the one that is used for Amrs. So luckily, having hers is hugely valuable because she's been through this before with a different type of robot. And so she's been working really hard, with, with the other folks on our safety team, like Kevin Reese, who's also phenomenally knowledgeable about the space to basically build out the standard and sort of get all the pieces in place. But that is definitely accelerating the process, because it can take a number of years for a safety standard to really go through all of its paces. But luckily, with some experienced folks on the team.

[01:01:56] Audrow Nash: It will be a bit. It should be.

[01:01:57] Pras Velagapudi: It should be easier. And we really hope to be able to get our vibes out in the world and get them, CE marked within the next couple of years.

[01:02:07] Audrow Nash: See, is like, what would a c?

[01:02:09] Pras Velagapudi: So C that's a marking that is used in, in the European side of things, to basically mark that you have approval under certain sets of standards.

[01:02:20] Audrow Nash: So for the standards so you'll make the robots, the robots will be capable of fulfilling the standard. The standard will be ratified at some point. Do you have like do you have like an optimistic and pessimistic estimate for the standard?

[01:02:36] Pras Velagapudi: I think it's it's probably, you know, I think I don't necessarily know exactly. Melanie is the person who has a much better intuition of it. But I mean, we're really talking, you know, single Digit year, couple of years, few years at most. We're really.

[01:02:55] Audrow Nash: It's like less than five, hopefully.

[01:02:57] Pras Velagapudi: Yeah, definitely. I mean, I'd say, you know, 1 to 3 is what I would hope. Right? Like, it might be quite quick, to get to an initial version of the standard and then we might. Yeah, you know, see where it goes from there or maybe, you know, takes a little bit longer, but we'll have enough of it sort of worked out that we'll have maybe a working draft or something that we can start getting in compliance with. But that's really more of her specialty than than mine.

[01:03:21] Audrow Nash: Yeah. Okay. Yeah. I've had Melanie on the podcast a few times. Melanie. Awesome. Hearing her perspective on that would be very interesting, but super cool. I really like that you guys are working on the standard, do you? So with your knowledge of standards, I know that she's kind of like she's been leading, right? Like she has the experience with the Amr one, but for robots to enter the home, or say like a grocery store or places where there are more people that are less trained, how do you imagine that will occur? Do you imagine that we're going to wait on standards like this? Or will it be that companies just kind of go with it and suffer the legal consequences when they, like, step on a kid or something? Or how do you how do you think this space, like, how do you imagine the space moving forward?

[01:04:16] Pras Velagapudi: Yeah, that's a good question. I think that we should be building out standards as we go. I do think that once we have one good reference point, that does make it a lot easier to move to the next one in the next one, because we can largely build off of that or reference it. Right? It can leverage it. I would hope that we're not going down the route of people just winging it and trying to get insured anyway, and then having something horrible show up in the news about, a robot stepping on someone's toes or something. I don't want us to. I mean.

[01:04:48] Audrow Nash: That already happened in some places, but.

[01:04:50] Pras Velagapudi: It's definitely happened in some, at some conferences and trade shows where they've been showing off smaller humanoids doing stuff, because those humanoids largely, run without too much safety system. Yes. Oversight at all. For sure. But I don't want that to happen. I mean, I think it could be a big, black mark on the industry. If that's how we approach the problem. I'd really love to, get there in a safe way and with a pretty good understanding of what we're doing now. I think there's two pieces that might accelerate that. One is that obviously with an example reference of like, here's how we're doing it in industry X going to industry wide, it's a lot simpler like we said before. And then I think the other thing is if we really show a lot of capability, then there's a lot of industry pressure to to figure out the standards and that sort of like naturally, accelerates the development and the, the working groups and really the whole process, really if a lot of people want a thing, right, then it tends to move faster. And so I think we will see that hopefully that the capabilities that we're able to demonstrate in some sense, builds a lot of, goodwill and interest in building out safety frameworks that can support doing it, and maybe there'll be interim stuff that comes out of it, similar to how, drones initially had some experimental airworthiness certificates and things like that that let people try out things under special qualifications, or that, you know, we'll see, changes in the regulatory space to kind of like provide, ways to phase in the technology in earlier stages. Right? In controlled settings and things like that. We might see some of that and we might also just see that, hey, you know, because it's such a high demand that the standards for things like, retail industry, just move really quickly because there's so much interest in it that a lot of people will spend a lot of energy, working on it. Yeah. Moving it.

[01:06:50] Audrow Nash: Interesting. How do you feel like, you feel like there's significant energy and, desire for robots to be out there by these different industries? I know you guys are most focused on material handling. But do you have a feeling for these other industries, like, is there a lot of pull from other places?

[01:07:11] Pras Velagapudi: I think there's a lot of interest. Yes. I mean, we certainly see a lot of interest. We have folks coming to us all the time asking, can you use Digit over here? Can you use it in my facility?

[01:07:22] Audrow Nash: That's a spot.

[01:07:23] Pras Velagapudi: Oh yeah, it's great. It's great to be there. But unfortunately right now we have to focus on a lot of those folks, which is unfortunate. And we're hoping making a big list of folks that we'd like to come back to, you know, in a year or two where we could say, hey, yeah, actually, now we can go look at that for you. And I think that's, you know, that's part of it. There's there's no shortage of interest in the platform. It can it can do a lot and it can do a lot very easily. In easily in that sense, not, obviously there's work to get. It's capable robots, but like, it's it's yeah, it's not the same level of work as, as deploying another type of automation where you might really have to. Yeah, design what you're doing around it. You can kind of slot Digit into the process that you have.

[01:08:09] Audrow Nash: Interesting. What do you think? So kind of a tangent. But it seems like a lot of my interviews, it eventually ends up that companies are planning on doing this kind of thing. So if it sounds like you guys are having a ton of demand for things that you can't service yourself, any interest in, like a robot app store or any sort of things where it's like, you let developers make applications for your app for your platform to solve some of the problems that customers might have. How would you think around that kind of thing?

[01:08:44] Pras Velagapudi: So we are definitely going to be.

[01:08:47] Audrow Nash: Moving in that.

[01:08:48] Pras Velagapudi: Oh, we're already basically moving in that direction. So one of the things we launched, last year was Arc, which is our cloud management platform for the robots. And within arc, you can basically build out what we call workflows, which is essentially telling the robot how to do different types of work by graphically putting together building blocks. And those building blocks are essentially skills they might be like pick up a tote or look for something in the environment or move to a location. But they're basically drag and drop high level skills, and sometimes they're powered by AI models that are trained on things. Sometimes they're their own, sort of specialized motions like docking sequences or things like that. But we're designing that platform with the intent of being able to hand it off to integrators and customers in the future and let them edit the workflows that Digit does on their own. They're very high level. They're at the level that you might write a work instruction for a person doing the same task, which is like, go to this work cell and, you know, look for a tote over here, you know, read the barcode on it, figure out which of these three places it goes and move it over there. Right. You can drag and drop that flow together. Or maybe in the future you'll be able to use a generic AI and just type that request.

[01:10:03] Audrow Nash: In and it'll what you want to drag.

[01:10:04] Pras Velagapudi: The box the blocks together for you. Right. Super cool. But that's our that's 100% our intent. We don't want Agility employees to have to touch every deployed robot. We want to be able to ship you a kit that, you know, an installer can put in for you, and that you can configure on your own. That's our that's our goal. Right. And so, you know, we actually used to have, you know, this, this internal tagline like we're going to create the App Store for labor. But then now with AI, it's like, well, you don't really have to write the apps anymore. So now we're just a store for labor, you know, maybe, that's the intent, right? We don't we want you to be able to we want to democratize, you know, democratize the the automation of of labor activities. Right? We don't want people to have to do work, and we want people to have to spend time working on the automation either. Right. We want you to just get a get a piece of equipment, get a kit that you can set up that just does these things for you.

[01:10:58] Audrow Nash: Yeah, yeah. One of the things I've been thinking that I would imagine robotics will arrive at at some point, will be that you can have developers that are building robots or using robots for applications, but they're not building them themselves. Kind of like what you're describing. But one of the things that I imagine is in specific domains, you're going to find that Digit needs to interact with something or there's something that's really hard perception wise, that you guys haven't solved, because it's another thing that you run into frequently or some thorny algorithm or something. Are you adding developer support so you can have like a module that, has custom code written in it, or.

[01:11:44] Pras Velagapudi: How does all that work? Yeah. So that's something that we've been playing around. It's something that in.

[01:11:49] Audrow Nash: Turn that's hard to do.

[01:11:50] Pras Velagapudi: Yeah. Well it's it's I mean, I'm sure you're very familiar with all of the various attempts at like, low code. No code. Robot. Oh, yeah. That's right. And yeah, there's, there's always the thorny edge cases of like, here's the thing that isn't one of the building blocks that you offer, but I need you. Right. There's two things that are kind of interesting about it for us from our perspective. The first is that we certainly can do the route of, hey, we'll give you a little sandbox block. That is something that you write your own, low level API calls into at some point. Right? Right now, it's not super easy for people to do that because Digits, control system and so forth are relatively complex. But this is where the second piece comes in, which is now that we're leveraging AI models a lot more for Digits, low level skills and controls, that's actually changing the game of how you specify things in this category of new skills in particular. Now, a lot of what we're starting to explore are things like physically demonstrating how to do a task to Digit and then turning into a block. Right. And that's that's where it's actually quite different from anything we've ever seen in the past where, hey, if you don't have a block you like, you might not author it by using code anymore, right? Which is really interesting.

[01:13:05] Audrow Nash: What a trip. Yeah, just a learning from demonstration based approach, but it's hard to make them general. But there's a lot of good work going on in that space I think.

[01:13:15] Pras Velagapudi: Yeah. And then there's also the question of exactly how general you need to make it, because it can still be a building block in a workflow. Right?

[01:13:21] Audrow Nash: So yeah, if you take a bunch of actions that already exist and just compose them with AI that watched it say, yeah.

[01:13:28] Pras Velagapudi: Or you just use AI for one piece of of the actions, right? I don't need to use AI to like, figure out a new way to move from point A to point B, like that's already a skill I have, but maybe I'll use it for, you know, stacking a new type of toad or interacting with a new piece of machinery. I haven't seen that before, but I can demonstrate that and make a new skill. So that's all stuff that we're, exploring very aggressively. It's not in production. Right. But it is stuff that we're exploring very aggressively internally and working with folks like Nvidia and Google, using their latest models on seeing how we can do on this type of new pipeline for creating new skills. And if generalist models do start to work well, we're selling a robot solution. Right. So we're perfectly happy to be leveraging.

[01:14:14] Audrow Nash: You're very well positioned right.

[01:14:15] Pras Velagapudi: We can leverage any generalist models that come out because as long as they help us do something, we're able to use it as part of our skill set. But at the end of the day, someone has to ship you a robot in a box and have a way for you to get that up and running and connected to your warehouse management system or whatever orchestration platform you've got. And we're doing all of that. We're the person that you can call on the phone that'll ship you in your arm when yours breaks, or whatever you might need. Right?

[01:14:41] Audrow Nash: Yeah, yeah, yeah. It's a good moat to have the physical hardware for this kind of thing. Because there's one AI generalist solution, and then another one comes up and it's like they don't have any moat that you just switch from one to the next. But if you are actually making the hardware, that's a much stronger position to be in. And any AI general solution that comes along is like, oh, great, throw it on our thing. It's real easy. That's great. So hell yeah. How are you viewing all this AI activity? And I guess breaking it down into different areas. And then how do you think about it? Melanie when I talked to her, it was one of the early podcasts when I started this podcast, was thinking of it kind of like really good search. How are you thinking? And that's specific for the language models in this kind of thing. And I know that Transformers are really interesting and they give you a lot of, like, interesting abilities across, say, vision or all sorts of other domains. But how are you thinking of it?

[01:15:42] Pras Velagapudi: Yeah. So we've definitely, been growing and evolving our AI strategy internally, and we use now a lot more AI inside of the robot than, than we did even when you last talked to Melanie. Because the field is it's moving so rapidly and, and we're benefiting from that greatly. I think there's a couple of different layers to it. At least that's how we think about it, is there's a couple of different layers to the AI models that we employ. And that comes not from necessarily what the models can do, but in terms of the places that you use them. And what types of requirements there are around how you use them. So at the very lowest levels of the robot, there's the local idea of what am I doing with my arms and legs, right? Like, what do I need to do to not fall over? And what do I need to do to interact with the environment to exert forces on that? And that's a controller that needs to be very close to the robot and work at a very high rate, because that's the rate at which disturbances to that model or definitely happen rates that might be hundreds of hertz, right? That's running right on the robot at hundreds of Hertz, right inside of our control loops. Right? Yeah, a.

[01:16:54] Audrow Nash: Lot of times a.

[01:16:55] Pras Velagapudi: Second. Right. And so that is taking as inputs, you know, some idea of what am I trying to do with my environment. And then it's going to figure out what do I do with my arms and legs. So the layer above it, that's like a controls AI model. Then a layer above that is like a skills AI model, which is really thinking, what are you.

[01:17:11] Audrow Nash: Doing at the, at the so at that lowest level, what kind of AI thing are you doing? Like a lot of like I talked to Jonathan Hurst at some point and a lot of it was like really good application of existing control theory, if I remember correctly. But maybe that's not true any longer for these kinds of low level command or low level control loops.

[01:17:35] Pras Velagapudi: Yeah.

[01:17:35] Audrow Nash: So are you using either and what kind of things are you doing.

[01:17:38] Pras Velagapudi: So so we are using AI there. And and we do have, we do have a good model based controller. So like if you see the robot working in production, it's right now flipping between the two. It actually has a really good model predictive controller. But then if things get out of scope of that controller, it actually switches to a reinforcement learned controller that's more robust, that handles.

[01:18:01] Audrow Nash: One that you just perturb a lot and simulate the one.

[01:18:03] Pras Velagapudi: That's been trained for a month. Yeah. It's been trained against a huge variety of disturbances. And so it'll actually switch to that. But what we're seeing over time is that that RL model is covering this pretty good, more of what the model based controller was originally tasked with. And so we think that that's so cool. Likely to sort of subsume a lot of that functionality in the future. Now that RL model is it's a relatively small model, but it's one that's trained across just a huge variety of different environmental conditions, especially around things like ground contact forces and slip and so forth. And it also takes advantage of still knowing a lot about the robot. So we do have the advantage of because we build the robot, we can actually co-design the robot and the RL together. Like sometimes we can actually make some design decisions to say, hey, we're going to change this actuation in our next version of the robot because we want to make it so that it's easier to model in reinforcement learning. Or hey, like, we can go get the actuator parameters or change the control layer.

[01:19:06] Audrow Nash: Coupling is so cool.

[01:19:07] Pras Velagapudi: Yeah, it.

[01:19:08] Audrow Nash: Is very cool.

[01:19:09] Pras Velagapudi: And it leads to very good. All right. That sim to real type transfer. So we can do what's called zero shot transfer, which is we can train a controller and simulation, run it the first and then run it first time on the robot. Yeah, it's on the robot.

[01:19:22] Audrow Nash: And so that's super cool.

[01:19:23] Pras Velagapudi: That loop of being able to just say, yeah, robot RL, robot RL is really powerful and we think is one of the things that, you know, we're we're really excited about doing well here at Agility scale. Yeah. So that all is that first layer. Right.

[01:19:38] Audrow Nash: And I have another question there too. And that's so, so cool. So I was I was in a like a local locomotion lab at University of Michigan for a while, and we did some reinforcement learning around gaits for a bipedal robot. And, it was very hard to use reinforcement learning to converge on stable gaits that we could transfer to the robot. So we'd learn it in simulation. We could get it doing really cool stuff. And then it was very hard to take that set of parameters that we gained, and transfer that to the robot to actually make it work like it did in simulation. Why have we seen so we've and then when there was this sim to real gap that we talked about a lot in robotics, how have we advanced so far so fast that we can do really good training in simulation and transfer it to the robot for zero shot learning in this kind of thing? Is it just, we know to vary absolutely everything about the model and the contact forces and things like that, and that creates a robust policy or what have we done to make such a significant advancement.

[01:20:44] Pras Velagapudi: Yeah. So I think it's been basically recently two factors. One has been that there have been improvements and, and more compute.

[01:20:52] Audrow Nash: Well, that's that was the other one.

[01:20:53] Pras Velagapudi: Definitely part of it. So so one of one of them is that we have made some advancements in the techniques like we're still using like PPO optimization like it. That part has actually stayed relatively constant for these types of low level controllers. But there are some new techniques. And for PPO proximal policy transition. Yeah.

[01:21:11] Audrow Nash: I'm surprised I have that in my head still. Okay.

[01:21:13] Pras Velagapudi: You go I mean, it's a yeah, you probably know it better than I do if you're a guy. It's been a while. Lab. Right. But, the key thing there is that, like that core fundamental piece actually does work pretty well still, with other techniques like teacher student systems, and, and basically other things that we've learned about how to build a curriculum over time to learn about stuff like locomotion, because a lot of people have been exploring it in this intervening period and found what works and what doesn't. So that's definitely helped. And I think that, you know, a rich, academic, history there has really helped us figure out these are the techniques that work really well for doing dynamically stable legged locomotion. And we've been learning that this whole time on the other side, which is really where just the benefit of everybody pouring so much energy into AI has been the compute that we have available. And GPU accelerated simulation is really a game changer, because now you can also just say, well, let's also just run it like 100, a thousand more. Yeah, trials than we ever could before in like one at a time. Right? You can simulate really fast. You can simulate on really powerful hardware. And that means to your earlier point, when you're talking about how do I make sure that I've captured the parameter space of the real robot, you can just domain randomize over that space and make sure you've captured it somewhere in that cluster of parameters that you're, that you're covering. Right. And so domain randomization in conjunction with really powerful compute help solve for some of this. And there is some amount of just hey if you train it for longer you can get better results. And now training it for longer doesn't take as long. And so that's that's really kind of all come together to make it such that, it really is highly effective to build out, legged locomotion in a reinforcement learning environment now and just have, an RL system basically trained up from whole cloth, you know, how to move in an environment. And of course, you can always add in additional stuff on top of that. Like you can, do some amount of imitation learning on existing gates and things like that to seed the optimization. But once you get into that base of attraction, pure RL can really go a long way these days. And so that's, you.

[01:23:36] Audrow Nash: Know, climbing keep getting better.

[01:23:38] Pras Velagapudi: Yeah. And it's been super impressive. Some of the results that that, research labs and, and different parts of the industry have been able to show off with just really, really impressive dynamic motions.

[01:23:48] Audrow Nash: Yeah, I guess, yeah. It just seems like our whole industry has learned more compute better. And it like is very true, which is just crazy.

[01:23:57] Pras Velagapudi: I think we're, we're seeing the same trend that's hit many other, areas before us. Right. Which is just that, there's a lot of things that, that we were trying to solve, you know, by changing the problem. And it turns out that we could just change our sort of that our cluster size and solve it instead.

[01:24:18] Audrow Nash: Yeah. Very much. It's it's so exciting and also a little disheartening to me for some reason, which is kind of funny because it's like you learn all these advanced things and like I just simulate it more and it works fine. Well.

[01:24:29] Pras Velagapudi: So there's a balance there. There's there's definitely learning structurally about like the techniques that work and those that work at any scale and are really important. But then there's also the stuff where, you know, like I did this, you know, in my PhD at Carnegie Mellon, I was doing a bunch of simulation with multi-agent, I think it was like Dec Pomdps, which is partially observable Markov decision processes. And then, yeah, decentralized versions of those where they're spread across machines. But now, I don't know that I would have to do any of that because I could easily probably do that same optimization on a single machine, and therefore I wouldn't have had to do nearly as work, as I was doing to solve, these decentralized policies because I wouldn't need them to be decentralized. Really. So so there's definitely this, this evolution of, like, the techniques that that we needed to do back then, a large portion of what we were solving for was actually just trying to work around limitations, compute.

[01:25:22] Audrow Nash: Limitations.

[01:25:23] Pras Velagapudi: And distribution of information. And nowadays, some of those problems have been solved and some of them are still fundamental problems. And, and yeah, benefit from these.

[01:25:33] Audrow Nash: Very true. Yes. For sure. Okay. So higher blocks you're saying.

[01:25:38] Pras Velagapudi: Yeah. All right. Let's call block number one. So block number two above that is is the skill layer. So the controls block again is I am getting some input of like I need to modify my environment somehow. I need to exert forces or do something. Let me figure out how to do that at the joint level. A layer above that is what am I trying to do, like some task specification? And then what does that mean? I should be doing a my environment. So I need to pick up a tote. The totes over here maybe I like draw a mask over it in an image or something like that. Or I give it a pose or something. Now I need some layer that's again pretty close to the robot because it's got to take into account changing perception information, but maybe not as fast, right? So maybe 10Hz or 30Hz and it's going to take that and generate out a new what should I be doing in my environment? I should be putting my hands here. I should be grabbing it like this. I should be exerting these types of forces. I don't know all the way how that's going to end up being joint level controls. But that's what I'm trying to do with the world. So we find that that's like a useful layer to sort of separate out and be training is idea of a skill, which is some sort of task specification down to some sort of specification of what needs to happen in the environment.

[01:26:55] Audrow Nash: Yeah. Okay. And then how are you benefiting from AI with that.

[01:27:00] Pras Velagapudi: Well, that's an area where historically you would get someone to do. And we still have quite a few, very talented folks here who write out that specification in something like Python. It's maybe, hey, if you see the toad here, move to this location, push the hands against it, close the fingers, pick it up, and then that's, you know, the pickup totes primitive, right? And we had to hand off those blocks and you might make a block for picking up a tote and placing a tote and moving between areas. And, and human skilled humans would write those right and still do for a lot of the more complex tasks. The interesting piece with the newer AI technologies, especially around things like, behavior cloning, is that you can do learning from demonstrate or imitation learning or even, doing things like RL to solve for how to do that layer of the stack. How do I get this task done? By modifying something in my environment. And so that's new. And because that has always been, a large chunk of building out the capability of a robot, we're now in this space where, hey, the thing that was the most expensive sort of incremental cost of going into a new environment is now working a totally different way. Yeah, it's it's, operating under a totally different model.

[01:28:21] Audrow Nash: With the 10% effort instead of 90.

[01:28:23] Pras Velagapudi: Or something, like just the type of data that you collect, the type of, the type of person that can be collecting data. So if you're talking about collecting something like tally up data, right. You might be looking at someone who's a technician collecting that data rather than someone who is a software engineer. Right. And so that was, you know, definitely move more quickly in building out these skill sets. And then you're software engineers are more dealing with the infrastructure behind it rather than tuning each specific skill, that you have. Yeah. So that's a huge game changer. And I think that everyone is scrambling to figure out how well and what the right techniques are for this or that. The space. This is the really exciting, oh my gosh, this could really change. Open up a lot of things. Yeah. How robots are scaling into environments. And so we're really interested in this third parties that, you know, like physical intelligence and skilled are really interested in solving for just the model part of it. But everybody is scrambling to figure out how this is going to be a game changer.

[01:29:23] Audrow Nash: Do you think that so, I mean, one of the things that we've been hearing for the last, I don't know, 20 years or something is like data is gold for all of this. And it seems to be true with all the processes that require a lot of compute. Do you think that we could enter a data is cheap point of view when it's like, oh, you just need to show it once it it learns how to do it. And then for a lot of applications, it's like the companies that have been collecting an endless amount of data, they're not really sitting on a mountain of gold like they thought for this kind of thing. Do you think we could enter that? Or is it a silly idea? What are you thinking?

[01:30:03] Pras Velagapudi: I think it's somewhere in the middle. Maybe. So in order to get to models that do really well with a little bit of data, they tend to have to be pre-trained on a lot of data. And so you end up with this situation where depending on where those models are coming from and how those are commodity sized, you may end up in a space where a little data is all you need, or you may not end up in a space where you need to collect a lot of data to build out that model. That lets you then move from point to point with a much smaller amount of data. And I think one of the big differences here is that in physical intelligence or certain of physical in physical AI, as the terminology has, has become known. Yeah. Physical AI doesn't have this long standing corpus of really well curated data for for the types of robot platforms that we're creating today. If anything, the largest data sets are probably single arm overhead tabletop problems coming out of, academia or some of historically the largest data sets. And so now we're seeing a lot of folks scramble to build up these data sets and build models off of them. But it's pretty different because we can't just go and basically download the entirety of everything that's that's, on Wikipedia. Some folks are trying to figure out if maybe we can use some portion of the videos on YouTube. You know, maybe that's a large corpus we could unlock with the right transformation. But that's a right now, a huge, difference between where we are with LMS and Vlaams, that are, that are working in kind of just abstract human reasoning and human language to, systems and models that are dealing with, physical or.

[01:31:43] Audrow Nash: Physical.

[01:31:43] Pras Velagapudi: Space. Right. And there there's a lot of differences in the data sets that people have available and how well curated they are, and the types of, robots and sensors that have been used to collect the data and how that affects how well it works on on other systems and how well it transfers. So I think we're going to see some amount of perhaps commoditization of parts of models. There. So it may be the case that there become established players, that produce, good models that other people can basically fine tune to their applications with very small amounts of data. It could be the case that, a really big entity like an Nvidia, puts out an open source model that sort of levels the playing fields, in the space.

[01:32:28] Audrow Nash: That would be like, amazing. Yeah. And they're certainly.

[01:32:31] Pras Velagapudi: Working on it with, with, with, efforts like Groot. Right. That's, that is the intent of efforts like that is to basically try and produce, models that sort of level the playing field, because of course, for Nvidia, the more people that are playing, the better they're doing. Right? Because they're.

[01:32:45] Audrow Nash: They're they're still in trouble.

[01:32:47] Pras Velagapudi: Right? Exactly. So we could see that happen to the space. I don't think we really know exactly where it's going to land. But I think what we're what we're seeing is that at the very least, we we have the, we have the amount of data where we can get, single skills or like small, generalized skills in certain local areas to work really well and really reliably. And so we know that that at the very least seems tractable and that seems like it would work. And that alone is already a pretty big unlock.

[01:33:20] Audrow Nash: Oh yeah. Okay. So then above the skill level I suppose you have something like a planning level or something like that.

[01:33:25] Pras Velagapudi: Pretty much at that point you start getting, it's a, you know what you might call a cognition or reasoning level. And that's really where Vlaams and La Amazing are really just, they're, they're right there, this at this level, you're basically starting to think agent like I where your skill is the API that you're interacting with. And so here you're really just leveraging web scale models. You don't necessarily really need to build anything that's to, no.

[01:33:51] Audrow Nash: Specific really of questions on this one. This one was kind of clear.

[01:33:53] Pras Velagapudi: Yeah. Yeah. The really interesting part, though, that I find particularly exciting is that because the space is so exciting to so many people, you really see really large players, spending time making models that are specific to, physical systems. So, Gemini, for example, you know, Google has been training on, robot data sets, which makes VMs that, you know, their versions of, of Gemini that are actually able to answer, or better answer queries that, that pertain to things like, objects in the world. Right. So historically is really hard to ask a GLM for things like, hey, what's the object that's, you know, to the left of my hand? Or like, what would I need to do to go pick up this object in the back of the scene or behind the table and move it to close to me like it did really poorly around egocentric reasoning and, you know, relative reasoning about things in the scene. But now, because so many people are applying these films to physical applications, that's now a specific area of focus that, the providers are tuning towards and trying to collect data and. Correct.

[01:35:08] Audrow Nash: Oh, cool. Wow. Okay. Yeah. What a thing. It's amazing. All of this feels like it's moving so fast for all of this, where it's like, I have no idea where we'll be, like, even like you think back to 2020 and where the state of robotics was, because that to me feels like a big indicator in time, because of Covid and everything. It's like where we were then. It's like everything has like now we have this whole AI boom and everything, and now there's all this interest in humanoids, and it's just it's so interesting where everything is going.

[01:35:39] Pras Velagapudi: Yeah, I think the that's the most exciting part about all of it is that regardless of what the specific outcomes are or exactly what the robots of the future look like, what's increasingly clear is that we can have them. Yeah.

[01:35:54] Audrow Nash: Super cool. What are you so going back to Agility, where are you guys headed? I know it's it's hard because, like, okay, we think back five years and everything has changed. Where are you guys heading into the next, like, 2 to 5 years.

[01:36:09] Pras Velagapudi: So yeah.

[01:36:10] Audrow Nash: So for us, the standard and unstructured environment or less, I don't know, less work environments. But yeah. Where are you going otherwise.

[01:36:19] Pras Velagapudi: Well I think the biggest thing for us is, is building out our V5 platform to, to really be the platform that we can scale out of the work cell with. That's really what we think is a gating function to there being a lot of robots out there doing a lot of different tasks. We can already, you know, we already have a fleet management system. We can already deploy new over-the-air updates to the robot and give them new capabilities out in the field. They don't require a huge amount of customization for different specialty tasks currently. Right? We don't we don't do that. We only build one platform and it's called Digit, right. So from that perspective, you know, that's all ready to go. And so if we could just get these last few unlocks around things like the safety, around things like, you know, opening up our manipulation capabilities with the tool changer, we're really trying to unlock these last few pieces to kind of open the floodgates for all the stuff that we've been working on. That's almost. You ready to be unlocked, right. All of the.

[01:37:17] Audrow Nash: Building.

[01:37:17] Pras Velagapudi: So for us, that 2 to 5 year timeline is really is a scaling timeline. It's get these last few pieces unlocked and then just open the floodgates. And so we hope that's that's what our next 2 to 5 years look like. We hope that we are going to be out there in the world that Digits will be doing, lots of different tasks, that they'll be easy to deploy, that they'll be out working full shifts, in the places that we can deploy them now from regulatory standpoint. And then that will basically be the, the, the driving function for other industries to want it. Right. They'll see them. And that'll become kind of a fire that that gets lit to say, hey, I want these in, in my stores, in my facilities, in my use cases. And that'll help really drive the, the ecosystem to, to scale further.

[01:38:13] Audrow Nash: Yeah, it'll be a flywheel. It'll be spun up at that point. That'll be nuts. So wrapping up, what do you hope that our listeners and watchers take away from this? Or do you have any calls to action, like you guys are hiring or anything else that you'd like to, wrap up with? Yeah.

[01:38:32] Pras Velagapudi: So certainly, yeah. If you, if you see open roles, you know, we're always looking for for talent. We definitely, you know, if any listeners are interested, I encourage you to go take a look at our job postings. I would say if you are a humanoid robotics provider yourself, I would encourage you to understand safety and start, you know, working with the standards bodies that are that are working on that. You know, if you want to join our, our ISO working group, I believe it's 25785-1, you can get involved, get, informed at least. Right? Don't don't get stuck where you build a perfectly working robot and you can't legally deploy it anywhere. Like that's.

[01:39:18] Audrow Nash: That'd be crazy.

[01:39:19] Pras Velagapudi: Exactly. Like, don't don't end up there. Everybody should be up to date on on this stuff. Everybody should really understand more. I think about what it takes to safely deploy robots in in production and at scale, because I feel like it's something that, you know, I certainly didn't even think about when I was in academia. I didn't even know most of these standards existed. And it's been such a core theme of what I've had to do to take robots to production since then in my professional career. So I think I think that's if you're a roboticist and you don't know, what safety standards are or like which ones are applicable, like go read up on them, go learn about them more, because they could play a big role. If you ever try to build a robot as a product. Oh yeah.

[01:40:02] Audrow Nash: How do you learn up on those? Well, like, what would you recommend?

[01:40:06] Pras Velagapudi: Yeah, I mean, I would recommend, doing some reading on, a field called functional safety and also, maybe starting with. I mean, it's, there's a lot of good primers on safety for industrial robots is a field that's existed for a long time. And even just seeing what that looks like and what the standards are and what they mean for different types of robots is, is, I think, a useful thing just knowing that they exist and maybe just ask the question of, hey, I was thinking about using a robot in ECS environment. Like what standards apply in that environment? You know what? What would I have to certify my robot to, you know, even just starting that thought process? And then once you really get into it, there's, there's third parties that can help you with that. And there's certainly a lot more, knowledge that's available, but I think it's just something that's not necessarily on everyone's radar. Like, every robot should have an app that you can get to to stop the robot. Like if you don't have one, put one in and make sure that if the robot starts going haywire, you could actually get to it, right? Stuff like that. Right. So I think that's that's important and not always on everyone's radar until until something happens and then and then it's on their radar.

[01:41:21] Audrow Nash: Yes. Big, big on their radar.

[01:41:23] Pras Velagapudi: Yeah. But, I think, you know, this is an exciting time. I, you know, safety is a huge component of this. I also think that, you know, if there's still any AI naysayers out there, I all three of you, like, come on, guys, like, get on the train. Right. This is this is happening, right? It's not a question of if. It's just a question of how much and how fast. And really, you know, I came from the position of, of doing a lot of classical AI, some more like optimization and search type strategies in the past and over the past 3 or 4 years, I've really, had an eye opening experience of just understanding, you know, how much is different and how much, capability, and flexibility is available in the latest generation of AI, model building systems and, and simulation and training platforms. So if you're not thinking about that, that's also definitely something that's worth, trying to although that field move so fast that like every three months, like you have to relearn half of it, but I.

[01:42:29] Audrow Nash: Know.

[01:42:30] Pras Velagapudi: At least start to keep your finger on the pulse and don't assume that something won't get solved because that almost categorically end up being solved very soon after you say that.

[01:42:39] Audrow Nash: I know. And how would you keep your finger on the pulse of AI? Because that to me it's like, I don't know, like I don't really have a good way other than talking to great people.

[01:42:49] Pras Velagapudi: Yeah, well, I do think that's a core part of it is, is, you know, pick a few sources of folks that, you know, are good, knowledgeable, sources that can parse the information. So I actually one of the things that I do personally is, is one of, our, lead AI engineers, Chris Paxton, runs a really active, social media presence. And so I'm often following.

[01:43:13] Audrow Nash: He's excellent.

[01:43:13] Pras Velagapudi: Yeah. And I think, you know, following him, following, you know, Jim Fan at Nvidia and other folks like that. So that's how I get a lot of my news is by seeing what they're responding to and their, assessments of different work that's coming out. I rely on a lot on that to get kind of, a filtered feed, because you're right, it is a firehose. If you're just trying to look at.

[01:43:36] Audrow Nash: Every and there's so much hype and there's so much because I mean, like, there's a lot of promise, but there's definitely a lot of hype. Also. And so having great folks like Chris filter it out, seems like a good thing. All right. Hell yeah. Well, this has been a lot of fun. Really cool to hear what you guys are doing. And that sounds really promising.

[01:43:58] Pras Velagapudi: Thanks. I mean, this is just a really exciting time to be doing this type of robot, to be doing this type of work. So, you know, it is really awesome to just come into the lab and just Digits off picking groceries or putting Halloween candy in a basket, and we're just like, oh.

[01:44:14] Audrow Nash: That's so cool.

[01:44:14] Pras Velagapudi: That's just what we're doing today. Like, great.

[01:44:19] Audrow Nash: Super cool. Hell yeah. Alright. Thank you Pras.

[01:44:23] Pras Velagapudi: Yeah. Thanks so much for having me. Bye.

[01:44:25] Audrow Nash: Bye everyone.