Senior Vice President and GM of Automotive & Cloud Computing of Qualcomm Technologies, Inc., Qualcomm
NEURA Robotics Executive Talk | Nakul Duggal (Qualcomm) on Physical AI
Physical AI needs more than Vision-Language-Action models to work in the real world. In episode 2 of the NEURA Executive Talk, NEURA Robotics founder and CEO David Reger sits down with Nakul Duggal, EVP and Group GM at Qualcomm, to talk about reflex speed and why machines need to learn from the physical world, not the other way around.
They dig into why Vision-Language-Action models alone fall short, and how NEURA is solving reflex speed in cognitive robots operating in the real world.
The NEURA Executive Talk brings together the people shaping what comes next.
Follow NEURA Robotics:
Linked...
About Nakul Duggal
Nakul Duggal, Senior Vice President and General Manager of Automotive & Cloud Computing at Qualcomm, has been active in discussing the company's strategy around physical AI, automotive technology, and robotics. In a March 2026 interview with NEURA Robotics, Duggal said that the idea of automating what humans do in a heterogeneous physical environment is a "new concept" that is "starting to get explored very rapidly." He stated that "the future will not be decided on screens" but "built in the physical world through powerful partnerships." At Qualcomm's Investor Day in June 2026, Duggal described physical AI as "the next great computing wave" that runs on the edge, not in the cloud, and said the addressable market for automotive, industrial, and robotics is expected to grow from $300 billion to over a trillion dollars within the next decade, with robotics becoming a very large segment. In the automotive sector, Duggal announced a partnership with Stellantis in May 2026, stating that Qualcomm will deploy its digital chassis platform across all of Stellantis's brands starting in model year 2028, covering connectivity, in-cabin experience, and self-driving software. He noted that Qualcomm's automotive business has grown at about 25% annually and is expected to exit fiscal 2026 at $6 billion in annualized revenue. At the Qualcomm Auto Summit in June 2026, Duggal explained that the company decided three years ago to "overdimension the silicon" and build a mixed-critical architecture that allows cockpit and ADAS systems to run on the same platform, calling this approach "Flex." He also said the industry is moving from software-defined vehicles to AI-defined vehicles.
Transcript (18 segments)
N
Nakul Duggal0:03
The future will not be decided on screens. It will be built in the physical world through powerful partnerships. About this amazing moment also talking about our partnership and what we have in mind or what we were working on, describing it quickly and defining a new standard for physical AI. Things have come together this quickly. I cannot imagine that from the time that we first met face to face to here we are in 60 days and announcing this collaboration. I think it's a really interesting time for physical AI because I feel like AI is looking for the next big breakthrough for where the technology shift takes us. And I think over the last 5 years it's just been incredible in terms of how LLMs and VLMs and VLAs have expanded, but the idea that you can actually automate what humans do in a physical environment that is very heterogeneous, that's a new concept. So I do feel that this is one of those problem statements that are starting to get explored very rapidly, but there's a lot of new art in front of us. So that makes it so exciting.
D
David1:24
They're making physical AI really reactive, being able to understand the environment, but actually being able to react to different kinds of situations. And I think this is also the necessity of a new level of physical AI. So let me, the way that I've started to think about robotics is when you are trying to do something that has not been done before, you have to break it down into how will you learn something that has not been done before. And I think as AI has made it evident, there is a data loop, there is the accuracy and the quantum of data, how much data do you need to where you can actually train AI to understand what the rewards are, how the reward function works. But with physical AI, there is also the aspect of the sensor itself because in physical AI, you don't have any historical data. You've never really dealt with what it means to touch something, hold something, move something, assess what it is, and how do you deal with it? And that's a part I feel beyond the brain that actually has to do with the nervous system. It has to do with connecting things that typically we haven't ever had to go deal with. So I would love to get your thoughts on it because you've been working in this space for such a long time. That's the core, like it's not just to do things or react on things you see but actually beyond that. So it means when I'm taking this glass or this cup to figure out the shape of it, like how does it look beyond the cup which you don't see, and for that you do need to have a reactive nervous system to be able to really interact and react. For the autonomous and even more, I mean we are right now talking about mainly solving industrial challenges because we are living in a world of having the same labor shortage. And in that case, we saw that vision language action models was just simply not enough to do a real task a human can do, simply because there was some information missing and the other part is it was way like you needed too much data also. So it means we right now lack data. So if you're just using one approach like just the vision, you actually need much more data on video stream than you would need if you also use the other sensors we have. We want to solve that you can actually train a robot a new task in less than 3 hours because if not, it will not be scalable. And for that, I think the biggest requirement was how to gain the right data, at the same time how to train it in much faster way, and the third actually even how do we enable the world to be trainers instead of us doing the training for everyone else. I think in addition, I feel like this is something that I picked up in one of our previous conversations: when you're trying to automate something, the tendency is to think about how would a human do it, but a machine could do something quite differently than a human. So when you think about degrees of freedom...
N
Nakul Duggal4:47
Yes. You can think about degrees of freedom of the sensor but you can also think about degrees of freedom of the model itself. What might be the easiest way? If it's multimodal, like if I look at this cup and I don't see that it has a handle, how would I manipulate it and do I have the ability to figure out on the fly what to do?
D
David5:07
Exactly. So I feel like this is one of these phases where there is a new level of learning which is very reflexive as opposed to something that is pre-programmed or something that has a set of rules that are predefined that you have to deal with. And of course there have to be constraints because you have to design some kind of an actor where you know this is the maximum that I can get out of what the physical sensor is that I have, but then you also have to be able to think about what is the level of coarseness or precision that is required to complete the task. So if the task just has to be completed but best effort is good enough, what does that look like? Or if it is a very precise machine welding approach that has to be done exactly a certain way, what does that look like? So I feel like there is a very interesting kind of correlation between the automation path that we are on and how are we going to get there, what are going to be the steps. And I would love to get your thoughts because I'm sure you spend a lot of time thinking about what is a core skill and there are lots of core skills versus what are fine skills and what is the order in which you have to solve them and what are the dependencies, what are the interdependencies between the actor and the training and the environment, and perhaps we'll talk about portability of those skills.
So it's a super good point because how do we make physical AI be a global intelligence without actually being forced to have a certain embodiment? So it means, and it's very simple also to explain: if you're working only with vision, then it matters that you know exactly the kinematics of yourself because you do need to calculate everything to it. If you're actually working on the senses of also feeling, then it doesn't matter if my finger is thicker, you're still able to press it the same way and you just basically trained on force feedback or the sensor input instead of the kinematics itself. And this is also the necessity, I think in the future to build a physical AI system which is not depending just on the form factor of a humanoid or just on a certain humanoid form factor, but you can actually think physical AI over every physical device which has somehow sensor input. If we really believe that one company itself can solve all world's issues themselves, then I think the media would be right to say it will take too many years until this gets real. I think the possibility of having a perception language action model is enabling basically the whole world to contribute to make this real in very short time frames. First because you're shortening the training time, second because you're not caring a lot about the embodiment itself and the form factors of it. I really like the concept of Neura where your model is, of course we'll build our models but we don't have a problem if our customers and our partners also build their models because it really is about having access to the data, access to the embodiment, trying to figure out how to go solve these problems, and there is not necessarily just one single way to go solve these. You know, when you came up with the concept of the central nervous system, it makes a lot of sense because you are essentially trying to figure out how does the human body operate? How does it react? But when you think about converting that into machine intelligence, the nervous system intelligence is quite different from the brain intelligence.
N
Nakul Duggal8:47
Exactly. What kind of goes into that thought process where you say I need to have essentially the equivalent of a brain or a mind for the nervous system? And you know you have a lot of history in control systems and robotics prior to Neura, but give me a sense as to how did you guys come upon that thought process and what's the vision for an intelligent limb or an intelligent foot?
D
David9:13
Mainly because we came across a lot of boundaries and borders basically with just using visual language action models. If you're coming into the real world, into a real manufacturing space, you will actually find that most of the tasks humans are doing today are not vision based, they're mainly skill based, and the skills mostly based on more data input than just vision, but mostly actually the field of touch. And this is actually where we come from. Then we started to use just the feel of touch and then the vision, and then you will figure out like when you press something into each other, you can actually feel certain maybe false limits, but you still didn't hear the click where you know you reached actually the end position, things like that. And this is also where we simply figured out that even if you do it with vision and you do it with force, the most important thing is also then still not calculate everything in one brain. Having the knowledge in one brain is important. All right. So we know how to do things, but then by being reactive, you do need something like a nervous system, and this is where we get limited because you get slow. Mostly when you watch today humanoids also, you even do them fast, and mostly you do them fast when it's just a vision task where you drop things. So this is something we can do super fast, but if you are actually doing a real task, then you're going super super slow because also the way of compute and the model even itself has a delay, and the delay causes that you have to be slow to catch the delay and at least don't destroy anything. So here we came up, okay, let's do something like a nervous system, where you train also not just the brain but actually also this nervous system with a super fast, let's say smaller neural networks, basically super reactive, which are actually fulfilling the task while the task actually is given by the brain. So if I come here and I feel now that the cup, like now I see, but if this one would be closed and I wouldn't say there is water inside, it could be also without water. So what we do here is basically that we interact right now with our nervous system, and our nervous system is right now depending on the shear forces of the thing as the tips are. So it means I know how much I press on the glass without letting it fall. Even if I change now the wave inside, it's not calculated in the brain. It's actually just a reflex we have in our memory effect of the muscles and the reflex basically of the nervous system where we basically react on the change of behaviors, and that's how our real world looks like. It's never perfect. It's never not moving and static. It's actually everything we do is flexible and we have to react on situation, and that's the best ability a human have, why we are actually reacting on abnormal situations. And this is also where we came up with the core idea. This is something we also have to put into the physical AI brain and distribute it on two levels on the architecture to enable it.
N
Nakul Duggal12:40
You know, as you describe that, if I try to think about it in terms of what would it mean as an architecture? I have to think that it has to be like an AI control system because an AI brain we understand because that has more to do with reasoning. It has more to do with I saw something. I understand what it is. I go back into my memory bank. I now know what it is. Therefore, I need to figure out what to do next. But this whole concept of how do I grip? How do I move? How do I feel and then make a decision? You can't equate it to a brain because there isn't anything that you are reasoning about. You are actually delegating it to a control system. But the models for this have traditionally been very linear. They've always been rules based. They've always been, you know, I think like you it's a very good example. Hey, I hear the click and I know that the joint is in the position it needs to be. Now I can actually take action. I have a rule that I have programmed and therefore whenever those things are heard I can now take the action. If you start to move that away to something that is more generalized where rather than actually making it rules based you're basically saying here is the actor that I have, what are all of the things that I can do with this actor? What are your thoughts on the coarseness or the precision that is needed? Because in my mind for example, if you want to do something in a very precise way, the question is do I have to look? So do my eyes have to get involved as I'm trying to go do something versus if I'm doing something coarse, I'm pouring water to you. I can look at you and I can pour water because I know physically where things are. What is your thought process around what is the sophistication of the actor relative to the coarseness or the precision of the action that you have to take? It's mostly not the precision, it's basically the reactiveness and the reflex. Exactly. Like when you're even like pouring water inside another, why it works because you even hear and you feel like even in this moment how you do it and this is also how you react on it. Because if you wouldn't have the feel, you would maybe just over pour, it will just everything pour, you would even not be able to pour it in a very careful way because you might not see what's inside this carafe or whatever. So it is always the speed of the reflex and this is also why we are so much into this smart limb concept, simply because this is something you can train perfectly and while giving the input from the brain towards the smart limb which is right now responsible for the reflexes basically by doing a task, you can actually have the speed managed in a way which you need to react on this real physical world. And in the same time, the interesting thing is that you don't have as a user or somebody who really takes advantage of this concept, you don't have to think about it. And that's actually the most important topic. I think it's not getting more complicated now because simply what you do is by gaining the data out of data suits or out of the robot or even out of simulation, you're also having the same concept again that you're gaining there only the data which are required to do this fast reaction. So you only care about this little nuances basically of the nervous system to be able to fulfill the task properly.
When you introduce the concept of the kinematic model for the specific embodiment. So if there is a physical model that I know the embodiment is designed around, that provides a lot of rules in terms of okay I know what it is, the constraints of the hardware are based upon this, I can have an implementation that allows me to operate with a certain degrees of freedom. So depending upon what the task might be, I can use different types of modalities, I can have a different type of capability set, I can keep evolving. That concept to me was very interesting because I've heard about that concept in the context of locomotion. You know, we've seen videos where if you have a robot dog and you chop off a leg, it will still keep walking because it is essentially designed to reconfigure its kinematic model with one less degree of freedom. Now it knows what to do. But if you start to apply that to the next set, right, where you say, 'Hey, I can pick this cup with all five fingers. I can also pick it with two. One is probably not possible. I can pick a different shape.' That gives you those degrees of freedom. And I feel like that is probably something that is really interesting to dive deep into as far as what will PLA models be designed for, because you can add a lot of complexity. You can also make them very simple such that there are tasks that they must be able to do that are highly generalized and then you build capability on top.
How does that interplay with the types of problems you want to solve? Because there are so many different problems, so many different environments that require automation. How have you thought about how to segment? The smart limb concept is like human. You have a nervous system, you have a brain. But I think it's even going beyond human because the good thing is, and it's also where Qualcomm partnerships is very important, is interconnectivity actually with all the other surrounding information. So it means that in this world, being a robot, you don't have to go to your fridge and open it to see what's inside. You can actually interconnect with the eyes of the fridge if there is a camera inside. The same also with autonomous driving vehicles. You can interconnect with any other sensor around to actually give information and react on them which you're coming close to. And I think this is actually the next big step for us, is like going beyond human capabilities and saying, okay, you actually have also another angle to work on. So you have two robots interacting with each other. You could actually use robot number one can use the eyes of robot number two if this is a better angle to see beyond, to be able to react and solve the task. And this is also the next big step we have to do to bring artificial intelligent physical AI actually into the broader space, and in the same time also to bring them into industrial space, because it's all about interconnecting, it's all about talking to devices without maybe pressing a button or without just sending commands and having a response basically on that.
It's very interesting you say that because when you were talking about Neura AI and the ability to see what another robot sees or another, you know, say an appliance in the home, what is the oven scene, etc. We are working with a team inside Qualcomm and it's a camera AI team, and the technology that they've built is the ability to download a model into the camera and you can design the model to do whatever you wish it to do. And then it will have a bunch of triggers that allow you to then ask the camera to stream if you want more information. And then you can define what is the level of streaming capability you need: low definition, high definition, you want to zoom in on something. So you have this, it's almost like I call it a close caption stream, and I have early information on what the camera is seeing. And if I'm interested, I can start to say, I would like to know more. Please start to send me live data because I don't want to trust your model. I want to just see the stream and send it to a more capable model where I will process it myself. And when I was listening to you talk about robots connected into an intelligent environment, it is very similar because it's like I want to understand what is the physical state of something that I have to ultimately plan to engage with, and that creates this network of robots and the environment that the robots are in that is essentially like a vision language network because you are converting it to a language that the robot can understand.
Ultimately, it will potentially even go one step further. So, I feel like the deeper you go into automation, the more of these capabilities you start to think about in terms of how and the other piece that is fascinating to me is that almost all of this has to be deployed at the edge. Yes. Because this isn't something that you're going to go to the cloud for. The latencies don't permit it. So, you will either have something that is an on-prem deployment. In many cases, almost all of the inference will happen at the edge. Maybe there is a collaborative network that works that robots and local infrastructure works with each other. So to me it's really interesting space that we are starting to get into.
I think this is also the core thing is why we talk always about three level architecture. So the first is like nervous system, second is the brain and third is the extended brain. And the way of how we do also the extended brain is not just all the time interconnecting basically with the arch device and be extended by that but more because the good thing is about if you're actually able to train the network the way to make more let's say environmental and situational depending. Then you can also have the external brain mainly giving you information by the situation you're in. Like you're coming into a kitchen you don't expect a CNC machine so you're expecting to like a fridge and microwave and whatever and dishwasher, and if there is a CNC machine inside you're allowed to think a minute about it why and download the information you need, but normally even it gets faster than a minute today. So this is exactly how we see the future. So that's also why the neurogyms we building right now, there's training centers for the brain, physically AI brain basically, we mostly get it into context. So you mostly have even the environment look like a kitchen or like an industrial environment because I think this is what the robot or the artificial intelligence mainly cares for, to give you the right information in the right time and as fast as possible so you can actually use the whole capabilities you have.
Yeah. What I find really interesting about your concept of the neura gyms and the neura suits is you are trying to replicate an environment in industrial cell something that you want to automate and then extract as much of the data as possible so that you can start to think about how to build not just newer models but even newer sensors. Yes. Because you have to be able to kind of get as close as possible towards the most efficient way do this. In the training centers you're figuring also out that's I think also the advantage like you're training you don't need much data so less than two hours maybe and then you're training and the good thing in the gym is that you can right away deploy it and check how good it is, does it work, does it not work, what's the limitation to make it work. And this is also right now the main task Neura is focusing on: building the pipeline the way that you don't need too much data, that you basically can train without being an expert, that is making huge progress by also being efficient with data. This is the core like how do we build a pipeline to be as efficient that you don't need tons of data for a small task. I want to be able to as a developer get access to a lot of data to keep improving my model and I want to keep validating how close am I getting to where I would say the model is not working good enough. So it's a full loop. Synthetic data, real data, stream it, store it, process it, train it, deploy it, look for how far, how close it is to where you want to be, and repeat the process.
And I feel like we have a lot of different capabilities now in terms of being very surgical about what problems to go solve. So this specific idea of building a perception language action model where you have specific modalities that you need to use like for example we use edge impulse with autoencoders, you want to be able to take time series data and you want to be able to look for patterns on that something you can train for. So I feel like this controller for neural networks for reflexes requires a very different type of infrastructure in terms of what you need to be able to go train these things and it will be very interesting how we take our combined assets and keep evolving this capability. I think it will help us a lot because this is exactly the core thing we are working right on is exactly to build the pipeline where everyone without experience can click on and get the right information out to train the model. So this is I think one of the toughest thing to shorten the training time by a lot.
Let's talk about ecosystem. Yeah. You know, I've most of my career I've spent developing ecosystems, being part of them, evolving them, transforming them, building new ones, and it's a very satisfying experience because you can see the before and the after state. We did that with automotive, that's been a great success for us, and we are starting to focus on that with industrial and embedded, and then suddenly now robotics is here. Our approach with the ecosystem has been you really have to focus on what is the problem statement that if you unlock then there is tremendous value for everybody, for society, and if you pick some goal like that and then you are true to that goal, then people want to come work with you. So one thing that we did last year was to acquire Arduino because we felt that in the age of AI, in the age of how quickly technology is evolving, you have to be able to connect to appeal to developers who want to be able to have access to the best technology and be left alone to develop. We just announced here at Embedded World the Ventono Q which is our second product that we've announced jointly with Arduino which is based upon one of our IQ8 chips that is a highly capable chip. It has safety, it has lot of camera support. It has a lot of AI, a lot of processing, multimedia. It's a multimodel chip and I feel like what we need to do at this point in time is to put a lot of capabilities in the hands of the ecosystem, learn from the ecosystem and then evolve. And I know we're going to work with Arduino to figure out how can we also get a new Arduino partnership going. Well, what's your learning been in working with ecosystems? I think that's exactly also again the difference from us to others is that we are the only one not claiming we can train the model to solve every task in the world. And I think here is very important to mention especially in the industrial space, there are all these companies having this know-how, they know how to do it and that's also why they're existing and this isn't something they want just to give everyone out because it would not make sense, you will have only one company then. And we started to build the Neura platform we call, and this Neura platform is there to build the way that in a developer environment where everyone can train the models and everyone can actually add their information to it, in the same time they can also keep their privacy by not publishing everything. And I think this is where today when I watched Arduino basically talking about what they work on, what's their network or how do you say the community look like with more than 30 million developers, this amazed me the most I would say because I really see this is the way we want to go. And I think the combination of the approaches we have and also the way like what we can probably learn from here, I think it's amazing. It's exactly what we were looking for because I believe that the future will be created by all companies in the world working together on solving that instead of one company making everything somehow and owning everything themselves. I think even in a data game, you can have companies like Palantir but in the same time you can actually just let everyone build their own, be their own Palantir. And this is the approach we are going right now: having everyone give everyone the capability to connect a sensor with an AI model on the platform Neuraverse in their secured environment and deploying the models there by themselves and also taking use of everything which is already developed. And I think this is the great vision of Neura and what we want to do. So here I really looking forward to learn a lot with you building ecosystems. I mean we are just on it. And looking forward also to see what we can do together on that.
Thank you very much David. It's been fantastic in the very short period of time we've known each other. I feel like this was meant to be. So yes, really looking forward to the core Neura partnership also.
D
David31:30
Also. Thank you so much. A great honor.