“If AI is overhyped, automation is completely underhyped. As automation advances and becomes more precise and durable and solves new challenges, AI is augmenting these systems by adding flexibility and autonomy.”

While we’ve not quite moved past the hype curve, AI technology today is quickly reshaping how factories and warehouses operate. In this special episode of the Manufacturing Matters podcast, TECH B2B Marketing’s Jimmy Carroll is joined by Melonee Wise, chief product officer of software and AI at KUKA; Juan Aparicio, founder and CEO of Reshape Automation; and Adi Dalvi, vice president of sales at OSARO, to discusses how AI is adding value to automation today, as well as what to expect in the future and what issues still need to be addressed. Together, they explore how AI adds flexibility to automation systems, from pick-and-place to AI agents that support problem solving and beyond.

The panel also covers physical AI, the importance of simulation and digital twins, the growing importance of data collection and alignment, and safety hurdles for humanoids and autonomous mobile robots. Additional topics include AI’s impact on jobs, why practical and well-defined use cases are winning over flashy demos, and major trends to expect over the next few years.

.  

Jimmy Carroll: [00:00:00] Hi everybody. My name is Jimmy Carroll. I’m the vice president of operations at Tech B2B Marketing. And welcome to this special episode of the Manufacturing Matters podcast. I’m really excited to have an expert panel of guests here to discuss AI and robotics and automation in general. Down the line here we have Melonee Wise from KUKA, Juan Aparicio from Reshape, and Adi Dalvi from OSARO. Thank you everybody for joining. I really appreciate your time and I’m happy to have you. I’d to go down the line again, and if everyone could just introduce themselves, give their title, and explain what you do at your company. We’ll start with Melonee.

Melonee Wise: [00:01:12] Sure. I’m Melonee Wise. I’m now the chief product officer of software and AI at KUKA. At KUKA, I’m working on helping bringing AI together with a lot of the physical robots that we make, whether they’re mobile or traditional industrial arms.

Juan Aparicio: [00:01:31] And I am Juan Aparicio. I am the founder and CEO of Reshape. At Reshape, we do AI agents for industrial companies that specialize in all the back office stuff, like finding the right product, the right technical sales, customer support, things that traditionally have taken a lot of time and expertise. And so we built agentic application engineers and agentic customer service.

Adi Dalvi: [00:01:59] I’m Adi Davli. I’m the vice president of sales at OSARO. OSARO has developed 3D vision and machine learning software that we integrate with articulated arm robots for fulfillment applications, each picking applications, mixed depalletization, and some packout applications as well. So I’m happy to be here. And thanks for inviting me to join.

Jimmy Carroll: [00:02:23] Of course. Yeah. Thank you. So not too long ago I saw a really good executive roundtable discussion where they summed up the idea of 2025 was the year of uncertainty, and 2026 there’s a lot more room for optimism. And there’s obviously a lot of things to be excited about always. So I’d to just throw that question out there: What’s everyone most excited about in industrial automation right now?

Adi Dalvi: [00:02:48] Okay, I guess I can start. I think what I’m really excited about is pairing different technologies together to provide full solutions to our customers. For example, we’ve got a deployment that’s going on right now. It’s a mixed depalletization that uses two robots. And it also uses AMRs. So AMRs are delivering full pallets of materials to the robot. We’re picking off those pallets. The AMRs are also doing empty pallet exchanges, things that. So we’re not just providing one portion of the solution to our customer. We’re now pairing all these technologies together so that we can even have more impact on their ROI.

Juan Aparicio: [00:03:46] From my end, what I’m excited about is bringing practical AI, the unsexy side of AI technology, into this industry. There is a lot of talk on putting the applied or physical AI into the products, but then the back office still operates like almost 20 years ago in a lot of places. There is a lot of redundancy. A lot of friction. It still takes too long to get a quote out of some of these suppliers, and we want to help. I’ve been in this industry for 15 years, and I think finally the technology is at a point where we can move the needle and make this industry more efficient.

Melonee Wise: [00:04:31] And I’m really excited. And I think that this year is going to be the year that we start bridging the gap from these AI learning systems onto large-scale industrial automation. I think that we for the last couple of years, since it’s started, we’ve really had a lot of that connection from AI to industrial automation in research laboratories as cute demos and things like that. But I think that they’re starting to be a movement, and a lot of the work that I’m focused on is how do we bridge that gap? How do we build practical systems that make strong automation pipelines to basically make it easy if you come up with a model on your preferred platform, Gemini, ChatGPT, whatever the preferred model is, and transfer it directly onto these systems. I think that we don’t have a clear way of doing that yet, but I think that the work that I’m doing and that others are doing in this space is going to help make that possible. And I think by the end of this year, we’re going to see real practical platforms for doing that.

Jimmy Carroll: [00:05:44] Yeah. That’s a perfect transition to one of the questions I wanted to ask, which is this idea of different technologies, and Adi, you mentioned it as well, is different technologies working together to solve new problems for customers. And with AI, I know it’s a buzzword and it’s hyped and everything, and we work in marketing and we’re guilty of it as well. But AI, it’s everywhere now. It’s ubiquitous and it’s adding value in a number of different ways. And one way in particular I think is it adds some layer of flexibility to different automation systems. And beyond what you were just talking about, Melonee, what are some ways that you’ve seen AI adding flexibility on the factory floor or in the warehouse lately?

Juan Aparicio: [00:06:27] You want to take it?

Adi Dalvi: [00:06:28] Yeah, I can start. I think actually the executive panel, this was one of the questions that they also asked. And I think it was either Gabe or Mike Cicco had talked about, you know, kind of chaos in the warehouse. And that’s really where the machine learning or AI conversation comes into play. It’s being able to take these very unique SKUs and still be able to pick them with the gripper that’s on the end of their arm because the machine learning is really helping us do that. It’s not just vision anymore. So we in the warehouse, because we’re in fulfillment applications, we have to pick sometimes one of 250,000 different unique SKUs. And without that machine learning or AI component, we wouldn’t be able to do that.

Juan Aparicio: [00:07:31] And I think on one side AI is overhyped still, but automation is completely underhyped. What we can do with traditional automation, the level of precision and the level of durability and what we are able to produce is thanks in big part to the advances in automation, all the companies in this forum. So what I have seen be successful is wrapping these AI solutions, like bin picking, into a more traditional automation on the outside, like the safety side. You want to be more deterministic, you want to have your PLC logic around. But for some components, where you want to add flexibility, wrapping those around is what I see more, what I have seen working in real shop floors up to now. What could come next is to give more and more room to models where motions are learned, where it’s not just deterministic. When something changes, giving it more autonomy but always through boundaries. And that’s on the runtime thing. More interesting, being a roboticist, what is super exciting, what we have seen working from our end on the shop floor is a little bit less sexy but still very useful, is how that data that was in the software, how is that now consumed by other systems that were more on the IT world, like having agents that can create reports at 7 a.m. with all my alarms from an HMI. These kind of things that are more into the gen AI territories, becoming more mature and less risky. Now when there is risk, probably when you work where safety is involved and you have to move a little bit more cautiously, therefore things are a little bit slower.

Melonee Wise: [00:09:28] I think areas where I’ve seen a lot of success is, one, making the data more queryable. So large language models can ingest a lot of these large datasets, make it a lot easier to query, like, “I’m a support technician. How do I solve this problem?” So we’re seeing a lot of that coming to the shop floor. And I think that’s helping a lot of operators become a lot more independent because a lot of customers, end users, struggle with being tied to third-party support as opposed to having independence within their own ecosystem. And the other area where I see it is in the last couple of years, companies have started deploying their own lightweight models. They’re self-contained. They tackle very well-formed problems, let’s say in mobile robots, social navigation behaviors. You’ve seen that coming online a lot more with mobile robots. You’ve seen a lot more innovative stuff in self-contained problems like bin picking, like palletization, where those problems are well formed and well understood. And that’s been, I think, a really big improvement and it’s showing that we can industrialize these learning systems. But as I said, the gap is typically manufacturers or providers make those custom lightweight models. But it’s very hard for our end users to create something full cloth from the beginning and bring it onto their shop floor by themselves.

Jimmy Carroll: [00:11:05] The topic that was brought up a few times here is safety. And I guess one question I have, and I’m cheating a little bit because we’ve talked to a number of people here that have brought up robot safety, and especially when it comes to humanoids, but not just humanoids — with mobile robots too. Can AI agents help there? And if so, how? And beyond that, what would be needed for us to achieve the level of safety required for humanoids to be widespread?

Melonee Wise: [00:11:37] I have a lot of thoughts on this.

Adi Dalvi: [00:11:38] I feel this a perfect question for you.

Juan Aparicio: [00:11:40] Exactly.

Melonee Wise: [00:11:42] So I think that safety and these systems – right now we don’t have great methodologies for learned systems in the safety community. No one’s been using them as supervisory systems. I think that today the model has been, and I think for a while will continue to be, you have a safety island. Then you have non-safety systems sitting above that safety island. I think that there is an opportunity for potentially bringing that closer. I think that where you’re going to start seeing that first is in humanoids. The major problem for humanoids is twofold. One, don’t fall down and, two, know where all the people are. Because the meta problem is when you can walk around and touch everything, you now need to know whether you’re touching a human or not a human. And so where are you going to see some of the first innovation in that space is in safe human detection. And most likely it will take a multimodal sensory approach. So most likely RGBD, millimeter, wave, lots of different technologies, maybe even thermal or near IR to distinguish a person. And there are players that are just starting to get that out there. Neura has developed a ceiling-mount safe human presence detector, but it can’t be mounted on a moving ambulatory robot. And so I think you’re going to see safety innovation there. And we’re just starting to see those algorithms – because typically how they work is that you have a traditional algorithm and an AI algorithm paired together, and they’re a supervisory system against each other or voting system. And one can override the other to do this. And then on the other side is for don’t fall, one, doing models for doing kinematics of the robot and detecting when it might fall, figuring out how to bring the robot to the ground gracefully or avoiding those scenarios altogether and making sure that if there’s no humans nearby, you can just fall and not worry about how you bring the robot to the ground. But I think that’s where we’re going to see a lot of the introduction to AI methodologies in safety. But as someone who spent a lot of time working on standards, still we’ve got a long ways to go in the development of these standards, and there’s a lot of disagreement still on what it means to implement and safety-rate these algorithms.

Jimmy Carroll: [00:14:37] Sure. Could we get the robots to fall like those little toys when we were little? You push the bottom and they just crumple down. Any thoughts down there for you guys?

Juan Aparicio: [00:14:49] On the positive side, probably less runtime. Obviously AI introduces a lot of challenges, but they’re there for the opportunities that they can open. On the other side, what it could help is on all the safety assessments, all these preparations, there could be more help from the AI. And at least that could indirectly help, because a lot of times it is a lack of knowledge: “I just didn’t know that.” A lot of times it’s negligence. But if you make it so easy to the point that there is no excuse here, then I’m optimistic on that goal. And I agree with Melonee that this could be also on the runtime aspect of things. It is not a short-term journey here. Positive thing is, I can go to San Francisco and get into a Waymo and it drives me around autonomously. So there’s some precedents. I know that they are completely different environments, but there are some precedents there.

Adi Dalvi: [00:16:00] And for me, I think from a picking technology side, we do a little bit of this already. I think there’s some technologies out there that can sense if something’s in the cell that shouldn’t be in the cell. We do this for a mixed de-pal application, where we can train models, saying if this item or if this thing is on top of a pallet, don’t pick it. So we’re already seeing a little bit of that in this industry.

Juan Aparicio: [00:16:36] Very interesting. An anecdote wasn’t popping up in my mind. It’s so difficult to understand sometimes what is normal and what is not in a cell. I remember, many years ago, when we were doing our first bin picking — different company altogether, for a supermarket. And a supermarket you can pick anything, right? You can have bananas, and we were picking a carton of milk that was open and then everything was full of milk at the end. But the robot continued operating because we didn’t train for that. And now, in retrospect, if you can think, okay, if everything starts to be shiny and white, stop. You don’t want to be cleaning milk from robots. Otherwise it’s horrible.

Adi Dalvi: [00:17:19] We’re training things in open box detection. So if we see an open box, we have to stop. And we’re not going to go pick that thing, because we don’t want the contents to spill out.

Juan Aparicio: [00:17:33] Right. What I wonder is, if we need to anticipate all these cases. I can anticipate that something will be open or I can crash into a banana that is more soft than I thought. Or the other route is I know AI will be intelligent. Like a human, it will not need to be taught to do those things, or we will eventually arrive there.

Melonee Wise: [00:17:55] And I think what they’re pointing out is, although maybe it’s not going to help on strictly the safety case, it’s going to drastically improve the overall performance and deal with these extraordinary scenarios that make everyone’s life as an operator really bad.

Jimmy Carroll: [00:18:13] Yeah. Well, and maybe not a safety issue, but if the end-of-arm tooling goes down to pick something up covered in milk, eventually you’re going to have to replace that pick. You know it’s going to add to downtime. It’s going to be additional costs and things that. So maybe it’s not safety, but it’s operational efficiency. Macro-level question: We talked about AI and the hype behind it. Do you feel that we’re getting past the hype? If yes, what brought us there and if not, what do we we need to see to see that happen?

Juan Aparicio: [00:18:48] I always have completely two ideas in my mind that are fighting against each other. On one side, I think I see all this AI slop and, oh, we’re hitting a limit. “Why are you hallucinating now?” I thought that I’d tell a story to my kid at night. And it’s, wow, this is horrible. It’s not even well written. And on the other side, at the same time, I see videos absolutely auto-generated that are incredible. The advances of some of the companies, like Physical Intelligence. And it’s incredible. So I have these two ideas in my mind, or using cloud code and it’s incredible that you can do these things. It was only a dream before. So I have in my mind: Do we already have AGI or whatever it is, and it’s super useful. And we’ve just scratched the surface. And on the other side, understanding that this is not a magic wand that will be the answer to all the questions. What camp are you?

Melonee Wise: [00:19:48] I agree with you. I think what we’re actually learning is that when you know what you’re doing, you can really use it very powerfully and get really great outcomes. And so in that realm, we are finding what is possible post-hype. But at the same time, we are seeing this contradictory space of when you don’t understand how the tooling works or you don’t know how to get the right outcomes from it, the outcomes are very poor. But it’s incredible what you can get done today in terms of leveraging large language models, how much more efficient they make you, how easy it is to query these datasets. We’re starting to see a lot of value out of it. But I do agree that there’s this great U-shaped divide between the people who know how to use the tools and the people who do not.

Adi Dalvi: [00:20:48] We talk about this quite a bit, and we call AI, in our business it’s really machine learning.

Melonee Wise: [00:21:01] Yeah I agree. Absolutely.

Adi Dalvi: [00:21:03] Right. And so I think that’s the very unsexy way of talking about AI. But that’s really what it is. It’s machine learning. So I don’t think that the hype is over with AI, because as more people start to discover it, the conversation is going to continue to happen.

Melonee Wise: [00:21:24] I do think that we’re at this interesting inflection point where the barrier to entry is getting so low that we’re starting to see really super creative outcomes. So would you agree? All as technologists, we have a very specific view of how we might use the tool. And then we turn around and we see someone who is not in our field. And they use it in a way, and we’re like, “I never even contemplated the idea of using it that way.” And it’s stunning and it’s shocking and it’s really cool because you can learn a lot by seeing how non-native technologists use the tool. Yeah, it’s super cool.

Jimmy Carroll: [00:22:05] Well, I’m glad you brought up non-native technologists and sort of AI in general beyond this industry. I remember not long ago, I bought a portable fan to go to Disney with my family and brought it from an online retailer. And in the marketing, it’s like, “AI smart chip.” I’m like, “What does that mean? What is an AI smart chip?” But at the same time, you can’t ignore that AI is here and adding value in real ways now and making some things more flexible and easier to use. And the question here is, when you’re talking to somebody who’s not in this industry and you mention that you work in AI and there’s some trepidation there or even fear sometimes. How do you explain to people that you don’t need to fear AI?

Adi Dalvi: [00:22:56] I think it goes back to what Melonee says. It’s learning how to use it properly. It’s using it and understanding what the outcome is going to be. AI is not a big scary thing to us maybe because we’ve been working in this field for so long. But you know, going back to the hype question, I think the hype is now going to be around the pace at which AI continues to move. It’s just going to continue to get better and faster in a short amount of time. But, yeah, it’s not this big scary thing. You just need to know how to use it. And I think people that are worried about their jobs being replaced by AI, it’s not that. I think what’s going to end up happening is people who know how to use AI are going to be the ones . . .

Juan Aparicio: [00:23:59] For me, in my line of work, it really helps because at the end of the day, we sell AI agents, which we could sell also to healthcare or other places. But we decided, okay, this the industry that we know and the people that we want to help, but it helped me a lot to have sold automation before, because I ran into the same scenarios when I was selling automation to a factory, calculating the ROI, and what I always saw is that the operators that the robot was in theory replacing, they were moved to some other operations. So the retraining was important. Obviously some jobs were repurposed, but I haven’t seen any case where there were really layoffs because the robot started. And I truly believe in the mantra that more robots means more work. Same thing for AI for the back office stuff when we create an agentic sales engineer. It is not like companies that typically deploy it reduce the workforce. It is the same workforce. Now they can do much more than they used to.

Adi Dalvi: [00:25:02] It’s reallocation.

Juan Aparicio: [00:25:03] Reallocation. It’s kind of giving them superpowers. Now I can reply to 100 quotes at a pace that before I wouldn’t even contemplate.

Melonee Wise: [00:25:12] I think this a very nuanced topic. I think that technology has always had this struggle between policy and technology. Technology is always going to advance, but policymakers talk and figure out how we deal with the outcomes of technology. And when we look at what AI is doing, I think the biggest thing that we’re seeing right now, and it shows in some of the things that we’re starting to see, is that very senior people, large multipliers on their productivity, which right now is leading to less hiring of junior talent. And I think what we’re going to see over the next couple of years is potentially some challenges in how do we train junior talent to become senior talent, and how do we make sure that we’re building the right pipeline for that talent. But I think in the short term, we’re not going to see a lot of disruption. Also because we’re right now, in the United States and most of the world, operating at a labor deficit. So in the United States, in and manufacturing and logistics, were at 1.1 million open jobs. And so a lot of the automation technology, a lot of the learning technology that we’re deploying is making up for that gap.

Melonee Wise: [00:26:39] And how do we solve these problems? How do we deal with the displacement, because we all admit that there will be displacement in one form or another. How do we retrain the people to be able to do the new jobs that appear on the deployment of this technology? And that firmly falls in the policy bucket unfortunately. That’s not something that we as technologists are going to solve. And I like to say that if we invented an AI tomorrow that could solve 98% of all medical problems, a magical drug, I don’t think anyone’s going to be crying about “What about all the jobs for the doctors?” That becomes a policy problem. How do we incentivize doctors? How do we make sure that doctors don’t leave the field? How do we make sure that we still retain that talent and make sure for that 2% that we have a way of treating those medical cases? And I think that we’re starting on that kind of journey, let’s say, in technology for AI and automation.

Jimmy Carroll: [00:27:55] Well it’s such a good point. Especially look at some of the recent investments in new production and manufacturing facilities in the U.S. That gap or that ability or need to train people is going to grow significantly. And that next generation of workers, how do you get them properly trained? How do you get them interested in manufacturing? And obviously that’s an ongoing problem as well.

Juan Aparicio: [00:28:17] Electricians, for example. There’s so much need for them. To your point, policy will make a big difference here. We were pushed to certain degrees, and the AI has proven, okay, maybe there are other traits that become more useful and we need to incentivize.

Jimmy Carroll: [00:28:41] I love to ask this question of highly technical people, people I consider to be experts. I know there are certain terms out there that bother people. And maybe physical AI is one of them and maybe it’s not. I think most people know what people mean by physical AI, but what are your thoughts on the term? What does it mean to you? And for companies that are out there assessing whether or not any sort of physical AI might be for them, what do you recommend? Melonee, I know you have opinions.

Melonee Wise: [00:29:13] Yeah, I hate the term, because so much of that notion of physical AI is not physical AI. And when we talk about it or as laypeople talk about it, they don’t understand the nuance between safety and the physical controllers that we run on the machines and the real-time nature of some of these algorithms and stuff like that. I think what we’re actually talking about is application layer AI that runs on a physical entity.

Adi Dalvi: [00:29:57] I was just thinking about one of ways we describe what we do is we power the robots using AI. I feel like I’m always having to Google all these new terms that are coming to the surface, but it’s stuff that’s been done already or it has been done for years.

Melonee Wise: [00:30:18] Well. And it’s always sitting above the control layer typically. It’s very rare for you to learn all of the controls layer of the system using AI. Very rarely do you do that. And so it’s more at the application layer. And so I wish we would come up with a better way of talking about it, because when you talk about it like that — when you say it’s at the application layer, then the question is: Is it on a physical thing or is it in the cloud? And in some cases it can run in both places. And so by saying it’s physical AI, it’s like saying, well, is iOS on my phone “phone AI”? And when it runs on my MacBook, is it “MacBook AI”? It seems silly, and I find it kind of silly, but I understand why they do it. It’s to help people who are not experts get a better understanding that the tool runs on a physical thing versus quote unquote in the cloud. But even that is very gray these days. It’s very in the fog.

Juan Aparicio: [00:31:39] I remember still when I was working with Berkeley with Ken Goldberg, and they were talking about cloud robotics back in 2016. And then at some point we started running neural processing units very close to the inference, very close to the PLC. And I came to the conclusion, okay, we should call this, instead of cloud robotics, edge robotics, back then. But physical AI, it has a better rhythm to it, better marketing than just edge versus cloud. But I completely agree with the terms. On the other side, it also probably came into the marketing first because of Jensen and his presentation and put some light into this industry, into robotics. And then there is a lot of research, a lot of community building, more and more intelligence put into robots. But there is a positive thing of sending more light and more focus into this industry and more investment that probably some of these companies need or we all need. But yeah, it’s a double-edged sword too.

Melonee Wise: [00:32:45] It creates a lot of confusion with executives because they think it runs someplace special or it has some special association when it just adds more confusion when we as technologists try to set the record straight and help them understand what we’re actually building. And sometimes they’ll be like, “Oh, well, you don’t seem to understand.” And it’s like, “Oh, I have so many thoughts on that.”

Jimmy Carroll: [00:33:12] Well, that’s super funny. And I have another thing I want to bring up in a second that I think you might all hate, but it goes back to a conversation I remember having years ago with a brilliant guy. This was before the current discussion with cobot safety and everything. And he’s saying, they’re not collaborative robots. They’re force-limited robots. That’s what we should be calling them. And I’m like, “I defer to you. You’re the expert.” Kind of a similar idea. But within the AI space, last Automate, we interviewed somebody who works in agentic AI, and they said they have a number of companies that have gone to them because their C-suite is mandating the use of AI, which I understand it, but at the same time, for what? And I guess the “for what” is my question to all of you. How should companies go about assessing whether or not AI is for them? Where do they begin?

Juan Aparicio: [00:34:09] Yeah, going back probably to the keynote yesterday, it definitely is a trend and it’s a technology that is making things different than five years ago. ChatGPT, that’s a tool that I use every single day that I didn’t use before. And probably the only equivalent was when I touched the first iPhone, that sensation. The first time that I used ChatGPT, okay, I can see the trend. This is going to change the world, and that change is coming and it’s only going to accelerate. You cannot just turn your face towards the wall and say, “No, AI is not happening for me.” It’s going to touch every single business now. The problem will be then a lot of these executives say, “Okay, well, we use Copilot.” Then I check the box, right? So it’s also finding an ROI: What is the technology? What is the status on the maturity of the technology and what are the capabilities that we want to enable and say, okay, do we have a business case for this and work backwards, and it seems to be, okay, AI can be the technology for that. It’s also a bigger term for a lot of things, but hopefully it’s not just a checkbox saying, “I use Copilot. I’m fine.” That’s what we are also seeing a lot right now. We are fine, but does it really do what you want or not?

Adi Dalvi: [00:35:28] I think companies are just afraid of falling behind. And they’re just jumping on the trend. But I agree with what you’re saying is they need to understand if they implement some sort of AI within the company, what is it going to be used for? Is it going to be a productive use of their funds?

Melonee Wise: [00:35:55] I think the other thing is, all of what they said is true. But also the other thing is, yes, I think everyone needs to use AI. I think that there’s a clear value proposition there. But great, you’ve decided to go use AI. You have to make sure that you’re prepared to use it properly and your data policies, your semantic policies, and the action policies are aligned. So all of these systems, like LLMs or VLAs, they require certain things to really be useful. And so you have data that is consumable by these models. So first you have to get your data in order. Then you need to have semantics that are aligned. You can’t just go willy nilly and you don’t have your terminology or your APIs or whatever aligned in a way that is consumable and makes sense for the technology to use. And then if you are using a VLA or something that is meant to do actions in the world, your action language has to be well defined and consistent from system to system. So if you want a learning system to go do a whole bunch of actions for you, like go out and turn on a whole bunch of robots and make them do things, you’re going to want to make sure that all of those robots have the same action language, or else you’re going to be in this really deep hole of doing custom one-off integration for every piece of automation equipment you have.

Melonee Wise: [00:37:34] And so I would say that the big thing that I would recommend to executives who are looking at this problem, figuring out how to get into this world of AI, especially if they’re going into automation, is starting to ask their providers, how are you going to help me solve this problem? How do we align all the data? How do we make sure the semantics are the same? And how do we make sure if we are performing actions in the real world, that the action interfaces are the same? Because without that, you’ve got this AI that’s like, aha, I can’t do anything, because this wrench and this system calls it a quaternion or whatever. And, yes, they can make those assumptions in the background, but there’s a lot of niche assumptions that you have to make and can cause really poor outcomes. So that’s my little spiel on some of the gaps that we’re experiencing when wanting to deploy these systems in the real world.

Juan Aparicio: [00:38:38] And at the end, for some of the applications of AI, we are automating a specific task. And the same thing with automation. When we deploy automation in a factory, if your processes are inefficient, then the automation is going to highlight that. Everything that a human was overdoing just to compensate for the fact that the process was inefficient, now the AI is not going to do that. And then you’re going to highlight the inefficiencies. So then it comes back to: Do you have a partner that can help you? You understand that this is not once and done, that this a journey and that in parallel you will need to change things. You will need to make your process efficient. Otherwise it’s going to just be, oh, this doesn’t work. Because your processes are not the way that are meant. The same thing with the tools to be optimized by AI.

Jimmy Carroll: [00:39:35] I think in terms of assessing whether or not certain technologies might be a fit for a company, how important are technologies like digital twins and simulation today? I know these continue to grow, and obviously the Omniverse platform gets more and more popular. What are your thoughts on that and are companies doing this a lot? Is it growing?

Adi Dalvi: [00:39:58] It’s certainly growing. And I think more companies are doing it, but it doesn’t solve every problem. I think inherently you could build a digital twin that you think is complete, but when you implement an actual system on site, things are going to be different. The environment in production is always so different than with a digital twin or even a really good simulation. So I think it helps you maybe predict some scenarios and potentially helps you predict some outcomes, but it’s doesn’t solve all of the problems.

Juan Aparicio: [00:40:41] For me, the difficulty always was, there is the sim-to-real gap, and we can think about that. But there are also technologies bridging that. But the real-to-sim, the other way, it’s just very hard, like getting your full layout, everything that changes in real time, synchronized. So it doesn’t become a simulation, but really a digital twin is very hard. And then what you see is these two things that don’t converge and then you stop using it.

Melonee Wise: [00:41:09] I think the other thing is that we sometimes conflate a couple of terminologies. So there’s process simulation, which is: I can simulate how product will flow, let’s say through a warehouse, from conveyor to whatever. And I can put in some assumptions, like, the conveyor moves at this speed. And those are a little bit more deterministic, but it gives you an understanding of the process. Then there’s physics simulation. So I actually understand the inertial properties of the system. And sometimes we confuse those two. But typically process simulation is used for large numbers of things because it’s hard to simulate any one physics at scale, let’s say over thousands of pieces of equipment. And so typically you use physics simulation for “I want to understand how this one robot is going to work in this one work cell to do this thing.” And then you come into this problem of sim-to-real transfer. So, okay, I tried it all out in simulation. Now I want to transfer it to the robot. And there is a very big gap to that today. But I think that the first challenge that people have today is they don’t even understand that there are different types of simulation, and they are used at different points in the process, and they provide different types of value. But today, when we talk about the notion of a digital twin, people I think assume that we’re simulating all of the physics and all of the nature of the warehouse. But we’re not. And then the other thing that we don’t have today, but Omniverse is trying to solve that, and there are a couple simulators out there, which is reality sim, where you simulate the environment, so you simulate the actors walking around, causing chaotic energy in the warehouse. And so right now, we don’t have an ecosystem that brings that all together and seamlessly transitions that. Basically an engineer has to configure each one of those things and then use them at different stages of the life cycle of the project . . .

Jimmy Carroll: [00:43:26] Constantly tweak it.

Melonee Wise: [00:43:27] . . . to get answers along the way. And we’re just not at that point yet where you can seamlessly simulate each of those stages and answer those questions for yourself. And honestly that is potentially an area where AI could help us a lot, because it could help create those solutions for us maybe more easily. But that is a big gap that we have today.

Jimmy Carroll: [00:43:55] We’ve got a couple of minutes before the session lets out and we get run down here. But I do want to ask for everybody, if you have any fun or notable predictions for the next couple of years in terms of trends or technologies you expect to see in the market.

Adi Dalvi: [00:44:14] A lot of these technologies have matured quite a bit over the last couple of years. Call it AMRs, call it picking robots. I think you’re going to start seeing some massive scale within customer sites. I think from 2019-2020 to 2024-2025, a lot of end users, a lot of large end users, were running proof of concepts and pilots, and I’m starting to see these systems start to run really, really well, right? And I think the reason why these systems are now starting to run really well is because they’ve had that time to collect the data that was required for the models to be really good and work very well in these production environments. So I think you’re going to start to see some really material scale in these technologies at these end users.

Juan Aparicio: [00:45:30] I think we probably will come full circle to understand some of the things that Melonee was saying of the importance of semantics, the importance of data models on context graphs and things overlaid into agents as a multiplier. Giving it some understanding of how the world works and the relationships and the physics really helps systems perform much better than leaving them on their own to figure it out. And so that’s one of the things that we are starting to see, that creating these context graphs on top of a graph rack, this really multiplies the accuracy of these techniques and agents that will take longer, more autonomy, really thinking for a longer time, to the point that I think coding agents are going to get much, much better and continue to get better. Also videos. One of the things that I’m also very excited about: We are seeing that with images, going back to the photos that people used to pay for, making AI to do your photo shoot, and now Gemini can actually do a good job. But video is one of the things and also a longer horizon and really small clips of video completely auto-generated.

Melonee Wise: [00:46:53] I think over the next three to five years in automation, I think the thing that is going to be interesting is — humanoids are going to take longer than people expect — and in the intervening time, mobile manipulation is going to have a renaissance. I think that the VC community is going to start putting more dollars back into that space, and we’re going to see more mobile manipulation come to the forefront. We’re already starting to see that with some new startups. And I think it’s going to be interesting because it’s going to be the race of the technologies, because one of the advantages that mobile manipulation has over humanoid robots right now is safety, is the path to safety execution is much more well defined. It’s easier. And so I think that’s going to be a really interesting interplay in the next three to five years. And it’s really on the mobile manipulation companies and larger industrial providers to lose it, to be honest, because even from a cost perspective, they have the unique advantage of the cost is even lower. They have fewer actuators, the safety is cheaper. So I think that we’re starting to see it with VC investment into mobile manipulation, and we’re starting to see more of these companies be created. And so I think it’s going to be a really interesting five years of will practical solutions win out from hopeful aspiration?

Adi Dalvi: [00:48:43] Going back on that point, I’m really excited to see what technologies come out of the humanoids. The arms have more dexterity than just a traditional industrial robot. And could that be a technology that comes out of the humanoid that we can all now use to pick packages in the warehouse? So there are certain things that I’m excited to see that come out of the humanoid development that may be used in other parts of the industry.

Jimmy Carroll: [00:49:18] Or even outside of the industry. I was talking to Stu Shepherd yesterday, and he said one of the things that could be interesting with humanoids is seeing what developments might come in prosthetics for people. All super interesting. Well we’re going to get wrapped up here. But I would encourage everybody to go follow all these folks on LinkedIn and check out their websites. If everyone wants to go through real quick and just give your website:

Adi Dalvi: [00:49:45] So ours is just osaro.com, O-S-A-R-O. It’s actually an acronym. It sounds like a Japanese word, but it’s an acronym. It stands for “observe, state, action, reward, observe,” which is how we’ve developed our ML and AI.

Juan Aparicio: [00:50:11] I didn’t know that. I always learn something new. For us, it’s ReshapeX.com or ReshapeX.AI.

Melonee Wise: [00:50:20] And for me it’s kuka.com. It’s a large industrial manufacturer that’s been around for more than 100 years.

Jimmy Carroll: [00:50:26] I think everyone should know KUKA for sure. Well, I appreciate it. I really thank you all so much for taking the time. If anybody has any questions or comments, I would be happy to pass them along. You can reach out to us at manufacturing-matters.com. And once again, thank you all so much.