In this episode of "Visions: A Machine Vision and Automation Solutions Podcast", host Jim Tatum welcomes back Dijam Panigrahi, co-founder and COO of GridRaster, a company that specializes in next generation spatial AI and extended reality technology, who discusses AI's growing role in manufacturing—especially machine vision inspection—and presents a three-tier framework (human-in-the-loop, limited autonomy, full autonomy) for deciding when AI should act and when humans must intervene.

Related: Harnessing VLMs for Real-Time Factory Decision Making

Related: Vision Models in Manufacturing: Part 2

Visions: A Machine Vision and Automation Solutions Podcast, is the podcast for engineers, designers, integrators, and end users who want to keep an informed eye on the imaging and machine vision industry. Every second and fourth Tuesday we will explore the latest in imaging trends, developments and solutions. Here you will find interesting, useful insights and observations from expert interviews, solo episodes, even the occasional panel discussion, all of which aim to expand your knowledge on imaging and machine vision. 

Be sure to subscribe to Visions: A Machine Vision and Automation Solutions podcast on podbean, or wherever you find quality podcasts.

Transcript

Well, hello and welcome to visions, a machine vision and Automation Solutions podcast. I'm your host, Jim Tatum, senior editor of Vision Systems design and visions is an endeavor business media production from your friends at Vision Systems Design. Here you'll find the latest on everything from end user machine vision solutions to trends, developments, and perspectives on all things machine vision and imaging. Whether you've been working in the industry for a while or you're just starting to take a closer look at it, this podcast is designed to grow your knowledge and bring greater focus to your understanding of the imaging and machine vision industry. And now on to our show. 

Well, hi everybody, and welcome to visions. Today we're going to delve into AI's expanding role in manufacturing as AI takes on a bigger role in such applications as machine vision inspection, for example. One question keeps coming up. When should the system make the call and when should a human step in? Today we are pleased to welcome back Dijam Panigrahi, co-founder and CEO of GridRaster, a company that specializes in next generation spatial AI and extended reality technology. Dijan discusses a three-tiered approach to this balancing act between autonomous decision making, inspection, reliability, regulatory compliance, and operator trust, and why the future of AI on the factory floor may well depend as much on governance as technology. Welcome back Dijam. Always a pleasure to have you here. But if you could first off just kind of give me a brief overview describing the three-tier framework and how you came to it. 

No, this, uh, this, this has actually evolved, um, over quite a bit of time. And we can see that struggle, like with our various customers that we are working both on the industrial manufacturing and also defense, right. Um, so it's always been a case. What is that? You can allow an AI to make decisions and move ahead on its own, make, you know, and what is that? Okay. I want the human to be in the loop, right? Many, many, many decisions where you're looking at human ultimately kind of signing off or making the final decisions in many ways, right? And there are, there are certain processes which is off limits. Like, you know, AI is not at all trusted there, right? And interestingly, you could see some of this kind of play out very openly as well. Like you saw the recent disagreement between anthropic and the DoD in many ways kind of falls into that bracket where anthropic, I mean, they on one side kind of agree what they can be used for, but they somehow not comfortable taking the AI, which kind of does your maybe the automation of your supply chain and other stuff. But when it goes to the making real battle decisions, you know, that's where they are not comfortable, right? So, you know, we'll see more of this kind of play out over a time period. But fundamentally for us, right, because we have a vision systems, because we are in many ways, um, bringing the AI from like the back office to actual real environment, which consequences with the robots operating with operator around with real implications of something not going right. There is real implications there. So largely our play, and I think a large chunk of our customers are more comfortable on the tier where you have the human in the loop, you know, there are certain key decisions where the human have to make a decision, right? But 80% of the work could be done by AI. But there is the human in the loop, right? Um, but many of the I completely understand that many of the back-offices stuff that is kind of happening, like in terms of recording a data, you know, taking some recommendation out of it and then, you know, doing it's more on the back office kind of a thing. We see a lot of the fully autonomous AI agents kind of operating. So that's the spectrum, uh, which based on which, you know, this whole three tier and we understand why it is there, right?

Okay. So to follow that path, say, in a machine vision inspection environment, what kind of decisions should AI be allowed to make autonomously and which shouldn't, which should always require humans? Yeah. Aerospace. Yeah. Specifically. 

Sure, sure. I, actually, what has happened? Like, I mean, um, when you look at like, let's take the case of we are doing an inspection of a wing of an aircraft. Just take just the example of that, right? You or even a flight line when there is inspection that is happening, we you after that, you're making a decision. The aircraft going to kind of take off, right? So the moment I say that, I let the AI make all the decisions in a way, I'm kind of confident that my AI will do it one hundred percent right. So, so now the question is, do we believe that it has reached that level, that it can get everything 100% right? Right. Because end of the day, you are kind of you're training your AI models to be able to make decisions. But is there scenarios where the AI model may not be kind of, uh, exposed? Right. And and this is where the, the uniqueness of in human being there to see, uh, certain deviations and, you know, flag certain things, which maybe AI is not yet exposed to and may not be able to make those decisions. So when it comes to those scenarios, what, uh, like, for example, like when I am kind of making a determination during the inspection that what kind of and repair that a particular defect requires, I may use a AI agent to go and basically I take a picture of it and my agent just like, okay, I, it can measure correctly. Yes, I am confident it will do it. It can localize in the X, Y, Z coordinate where it is and all of that. What's the concentration of the various defects that are there and make a decision on, you know, maybe classifying them to be what defect type that we need to do. And then basically allowing, let's suppose a robot or somebody needs to act on that and activating the robot to repair those kind of defects. The moment you bring in that translation tool, providing that to a robot, or even providing the judgment to the human, or in some way that it needs to be a verification that happens by the human. And then there's another scenario. And that's why, like if you if you're aware, like FAA even today doesn't allow a fully autonomous AI in space, they do not agree. They it is still from even. I mean, there are many elements to it. Why why the AI governance is this way and there is a very good reason for it. I mean, for example, from a regulatory angle, they haven't seen where AI could be all encompassing and be able to do it all – no, they haven't seen that. And so finally, even for your aircraft to take off the human has to sign off on that, which means that you could in many ways, you're using AI as a supporting tool. Like, but that support is a big support, you know is maybe does your 95% of the job. But finally, that five percent still has to be done by the human to get this confidence right. So then, so you have a regulatory element that is kind of playing into the role of the governance of what is feasible, not feasible. And particularly challenging becomes an inspection is still, you know, it's very clear cut. And I don't know if you're, if you're tracking what's happening in different parts of the world, but if you look at, I think in general in Europe, they recently put some deadlines around. I think it was maybe here sometime, August, or something like, in 2026. I think it is August that, okay, before you can have this AI agents kind of play out in those operational manufacturing environments, you know, there has to be a foolproof study with the risk assessments and all of that to be done before you can use it. In fact, inspection has been kind of called out as one of the use case, which is considered a high-risk inspection, quality control and all of that. And they have to prove that with the data to be able to use the agentic eyes in a good way. So, and for us, and I would touch upon one other thing is, uh, in general, if you look at, uh, the industrial or manufacturing environments, um, which is very different from like a software and all like, and it's something that we have been able to better understand over the last two or three years as we kind of push this AI more onto the factory floors is you see how the systems have been built out because from an effective AI governance point of view, you it's not that we fit in something outside, but you are trying to build it into the overall system to be able to really be, uh, meaningful. We define the guardrails and, but those things has to be defined. There are very. So many system interplays happen, and you may have different agents trying to do different things, and all of a sudden you have so many agents that you're trying to kind of manage and it just becomes way too much. I don't think we have handled those many with that degree of confidence, but one of the biggest play that we have seen in those environments has been there is, uh, usually like if I have to broadly put, there are one that are operational, uh, operationally kind of focused technology elements. And there is one, which is more what they would normally call it, IT, the ERP and other systems that are there in all these places. Right. Um, so there is an interplay that happens right between the operational setup and the ERP system right now. If you see, and if you actually dig a little deeper, all the system operational systems, okay, the way the IT is integrated with the operational system is IT is almost like a reporting tool. There is not a bi-directional operational exchange that really kind of happening in those environments, right? In order to be really allow the agentic eyes like to make decisions and all that, even if you define the guardrails, then many of those data points that comes from the ERP systems or the IT infrastructure overall needs to be taken into consideration when you are making some decisions to automate, and I don't think those infrastructure exists today, right? So there is a infrastructural level kind of work that needs to be done to be getting prepared for like better even the AI governance, right? It's a very interesting intrigue. I mean, there are different pressures that are going on, but we absolutely see the immense value of it. For example, for us, for example, like, um, you want to kind of plan a robot to kind of do a repair of a certain object, right? Okay. And, um, you know, and it's a mobile robot. You're trying to allow them to do a path planning of it. Right. And we have now reached a, reached a scenario where I can just put on a headset without any coding, no path planning, anything needed. You just wear a headset, walk around it and you know, you say, okay, this is what it is and this is my target, okay? And you can let the robot then figure out everything on its own, right? We come to those stages, but there is inherent risks that are there. That's why we are forced to have human in the loop, making some of the final decisions, uh, around that. Right. Well, over time, I think it's also a question. I think the speed at which AI is going. I don't think we are able to kind of assimilate and keep up with it. And I think in general, like when you have humans working in those environments, firstly, you have to develop trust with the system, and usually the trust with the system comes over a time period of use. And I'm sure it's a matter of time we'll begin to kind of develop those kind of trust. And I'm sure the AI outcomes would. Um, and which AI outcomes are already validating anyway, but just that we need to get comfortable. 

Okay. It looks like you've at one point said that too much oversight can be as damaging as too little oversight. Uh, yeah. That being the case, how does a manufacturer avoid alert fatigue when AI driven inspection systems are generating review requests on the production line? Mhm. How how do you deal with that? 

Yeah. I think that's a real problem that we see too. I think, um, in many of the systems actually, we have um, become the bottlenecks, the humans or the operators that they are, they're kind of in the bottlenecks. We are not able to keep up with what AI is able to provide us. I think in that, like the way we see it, right? We believe that there are certain things. I don't think so. Human should be reviewing it. I think AI has gotten that good at that. Like this whole three. Yeah, this three tier thing that I'm talking about now, the three-tier approach has been kind of designed from angle right now, which is okay. What we have conceived that, okay, this is the reason, right? But if you break it down, even within that, a human in the loop kind of processes in which we are still having at many of the places the human to make decisions, I think over a time period, you will see that some of those processes, not all, maybe some of them, we will develop that kind of trust in the system to say that part doesn't need any more human intervention. I think we will reach that stage where there is a harmony between how AI agents are kind of working and the human in the loop, but it will happen over a time period when we have these data points proving that out over a really significant data that, okay, it proves certain things, doesn't still need the human intervention. And that's where I believe the pressure, also, the fatigue that you talk about from an operator angle will significantly reduce, but there is indeed a very palpable friction that exists today. 

Okay, okay. Um, ultimately, who should be responsible for setting escalation thresholds in an AI enabled vision system? 

Yes. Uh, we do. And, and the thing is, again, uh, many of those, uh, um, escalation, uh, that, you know, we, we have to kind of design into the system is still kind of in works like who is going to be respon, who is going to respond to it? As I said, there is a preparedness that needs on the organizational side as well. And when you're working within the regulatory framework and the legal framework, I think we did not talk about the legal aspect of it even today. The. How the --everything is defined, right? The ultimately, the legal consequences kind of lies with whoever the manufacturer or whoever it is, right. If with any of this AI elements not working out properly, I mean, the legal consequences still kind of stays with the manufacturer. I think there is somewhere I think those are the frameworks that needs to work that out. And I can understand where, you know, even with all the right intent, where is the compliance and the legal frameworks and all kind of playing a role into what can you do or what you cannot, even if the AI systems are ready to do it. But I don't think so. Our overall, how we run the business with all the compliance and frameworks, sorry, compliance, regulatory elements and legal frameworks and all of those things. I don't think so that those have evolved yet. We would need all of that to evolve through the process. And there is obviously efforts going on. There is an immense pressure from the business side of the thing, right, which is happening, but in a way, you're kind of managing or balancing the intent of a business to generate profit versus making sure things from a safety angle.

Well, to move on further with that, as agentic AI becomes more common in factory automation, uh, I don't know exactly where it is yet, um, but do you see machine vision systems evolving from defection detection tools into systems that can recommend or initiate corrective actions on the line? 

It is, it is, it is. Yes, we are there. As I said, many of the things we are kind of yeah, many of this thing we are already like where I was talking of the robotic path planning or talking about a robot going, and you know, we using those guys to do make the inspection of different places and just kind of move along. And, um, but considering, I mean, one of the considerations that is taken in, in the, in the space that we work in aerospace and defense is the value of the asset itself is so high, right? You know, they just cannot they're just not comfortable leaving it to any autonomous things, doing it. But I think when the same things are extended to maybe some structures and all, which may not be that expensive, I think they would be willing to maybe take that direction. Uh, but yeah, all, all those things are kind of playing out. 

Okay. So to get to the point where you have all the guardrails and all the things you're talking about in place, what kind of timeline are we looking at? How long do you think before that'll actually happen? 

Let me, uh, first, you know, see the position here, right? Right. Now, if you look at like, I think I was going through this report from Deloitte, I think they had an extended report. They were basically saying, okay, they had a study around that in terms of like, how many from industry point of view, how many are kind of adopted already, the AI and those autonomous pathways they're kind of taking and compare that with the AI governance elements. So it seems like there is a significant gap, which is very understandable, uh, that like, I think 74% of the enterprises have adopted AI, like AI and all of that. But when it comes to the governance angle, there is only 21%, right, from where they feel okay, they have something from a maturity level that they can do those governances, uh, make those governances happen, right? So you can already see the pressure is already there, right? You know, the advancements that is happening on the AI is going so much faster and there is indeed a business need, right? Ultimately, it's all driven by the business need, right? And so there's a lot of catch up that needs to be done from the governance side. I think that pressure is kind of happening. And at different levels. I think people have chosen different pathways. Like I think in U.S., we are seeing even it's driven from the government that we need to have the speed from you angle. They are saying, oh no, we have to be very certain about things before we bring in all these genetic elements. And then I think there was some really good stuff happening in Singapore, right? I think they have some, what is that called IMDA or something like that where the agency is driving. Okay. They know that nothing can be achieved to the perfection right from the beginning. So they have defined a model where they're going to iteratively kind of evolve this governance model, knowing very well that not everything is addressed right from the beginning. So, uh, very, very interesting how different approaches being taken at different places. We, we'll see what, uh, outcome, um, which one does the best? Uh, yeah, I actually cannot even say which one, which side should you take? 

Right. Yeah. Um, okay. How does the perceived pause escalate framework help determine when the vision system should trust its own judgment? 

Yes, I think. Let me give an example from an inspection standpoint. Right. So there are many inspections. Okay. So where I'll say based on what we are doing, we believe the vision system in many things can you can trust today. But again, I know where, um, some of those compliance or regulation things come as well. But for example, many of the defects, which is kind of innocuous, right? Or innocuous in the sense maybe it doesn't have a life-threatening implication. I'm coming from that angle, right. But those inspections still needs to be done. For example, you have a aircraft cabin where you are going to use your vision system to take a just use a headset, scan the, the, the cabin, and just see if everything on the seat is fine or if there is some tear or something needs to be done. Now, the vision systems can do all of those things, right? That's something even if you are able to completely offload it to the vision systems and say, okay, I'll take that one hundred percent decision of it and move ahead. Right. So what would happen if even if something is missed is maybe you have a dissatisfied customer, that's not an end-all, right. But when I come when I do like the rivets and all, like, what are there on the aircraft? If I try to detect if there is a rivet which is missing or a bolt that is missing, but is it something that you can completely depend on the AI? I would do a very good job. Like, you know, we can say to the confidence that almost 99 point, maybe 9 percent but I will never be able to say 100% that it would capture everything. It has consequences. So that's why you want to always have the human in that loop, right? Okay. There will probably not be anytime soon an utterly autonomous situation in that. Yeah. I think what, what would change is when you are doing the inspection of an autonomous system where humans are not flying, right? Because there's no that there may be a different conversation. 

Okay. Uh, do you have anything you want to add to what we've discussed? 

Uh, the only thing I would say is there's a real pressure coming from the business side of the value of AI that's been kind of proven. I think as we get more data points, right, you will see like, you know, um, uh, you know, if the operational issues that exist, so many of them and when you start using HDI, you say, okay, I got my maybe seventy percent of my operational issue resolved this year just using AI. It just builds those cases so strong, right? And it can do it. And we, we strongly believe it can do that. But I completely understand when you're balancing the, the business requirements, which is driven from a profit angle versus safety, which is driven from the human safety angle. I understand that, and I think somewhere those balances need to be kept. And I think it's a journey that we have to make. But we'll see more and more Agentic AI systems evolve. We get confidence with all guardrails that we're defining. The infrastructure is getting more ready to use, more data points to in real time make those decisions, because that's where the fully autonomous things will happen. And we can then trust the system decisions much better. 

Well, that's a wrap for this episode of visions, produced by Endeavor Business Media, a division of endeavor B2B. Thanks very much for tuning in. If you enjoyed today's show, be sure to subscribe to the podcast and share this episode with a colleague who would find it helpful. Until our next episode, you can find us at vision dash systems dot com or on LinkedIn, Facebook, or for more insights, updates, and breaking news to keep you in the know. Thanks for tuning in. Until next time, stay focused on your visions.