15 Jan 2026 · 51m · Michael Krigsman, Burkhard Boeckem · 19 quoted facts
Hexagon CTO Burkhard Boeckem argues that physical AI, meaning AI inside machines that move and act among people, must meet a far higher bar than chatbots. A hallucinating chatbot is annoying, but a misbehaving robot can injure or kill. His answer is to ground robots in accurate, constantly updated digital twins, design them to fail safely, and measure them on uptime and reliability rather than on flashy demos.
The central tension is between impressive demos and dependable work. Boeckem is an optimist, expecting autonomy in constrained settings such as inspection, picking and material handling within one to three years. But he warns that the hard parts are safety certification, maintenance at scale, regulation and enterprise readiness, not the technology. A general-purpose 'butler humanoid' is still well off, and boards that treat robots as a software rollout will fund endless pilots.
Boeckem says a hallucinating chatbot is annoying, but a humanoid that hallucinates or loses control can be dangerous. So the key metrics become predictability, uptime and safety rather than novelty.
A dimensionally accurate digital twin lets a robot make errors in simulation before it works in reality. It must be updated constantly, and the robot needs onboard sensors to notice changes like a moved chair.
Boeckem favors designs that separate the control loop from the AI and add redundancy. The robot should slow down or stop when it leaves the conditions it was trained for, rather than carrying on and risking a fall that hurts someone.
He estimates one to three years for autonomy in constrained settings like picking, placing, inspection and intralogistics. General-purpose, human-level dexterity in chaotic spaces is far off, partly because safety certification would be so hard.
Boeckem says boards overestimate the speed of deployment and underestimate the safety engineering, integration, service and maintenance involved. He says enterprises, not the technology, are the likelier point of failure, and that without a reliable digital twin a company risks funding a perpetual pilot.
Boeckem defines physical AI as AI embedded in machines like drones and robots that perceive, reason and act in the real world. Because mistakes can hurt people, it needs a much higher standard of reliability than a chatbot.
A digital twin is an exact, constantly updated digital copy of a real place, built from calibrated sensors and scans. Robots can practice and fail in it safely, which makes real-world deployment easier.
Boeckem calls it the big theme for 2026. It means controlling the whole robot system, including sensors, compute, controls and maintenance, so that it stays safe even as parts wear over time.
Boeckem estimates one to three years for autonomy in constrained environments such as picking, inspection and material handling. He expects small deployments on real factory sites in 2026, with wider rollout the year after.
Boeckem says boards overestimate how fast robots can be deployed and underestimate the engineering work. He says it is a complex system involving safety, operations changes, integration and maintenance, not a software rollout.
Stated by the speakers. Each quote matches the transcript, which shows what was said, not that it is true.
Burkhard Boeckem of Hexagon defines physical AI as AI that lets autonomous systems such as drones and robots perceive, reason and act in the real world.
“Physical AI lets autonomous systems like drones or robot perceive, understand, reason, and performs actions in the real world.”
Boeckem says physical AI should be grounded using calibrated sensors, verified maps and accurate reality capture to build an exact digital-twin replica of the real world.
“you use calibrated sensors, verified maps, do accurate reality capture in order to, uh, create a exact replica of the real world into digital twin.”
Boeckem says an accurate digital twin lets a robot be trained and make errors in simulation before it is deployed in the real world.
“actually train the robot in this digital twin and and try things out, make errors and and failures in the in the digital twin.”
When Hexagon launched its humanoid, it demonstrated the robot scanning a car door for defects with a laser scanner.
“our humanoid scanning, um, uh, with a with a laser scanner, uh, the um, the the door of a of a car for for defects.”
Boeckem says the best robot design has the robot recognize when it is lost and slow down or stop rather than continue.
“the robot does know when it does not know anymore, when when it's lost and slowing down, stopping”
Boeckem identifies functional safety as the big theme coming in 2026 for physical AI.
“In terms of of the I would say the big theme that is coming in 2026, which is functional safety.”
Hexagon's humanoid project built on its earlier experience with autonomous drones and quadruped robot dogs.
“our our humanoid project, and it was basically based off of our experience. We made with, uh, um, autonomous drones”
Boeckem says Nvidia is a long-standing partner of Hexagon in physical AI.
“as Nvidia is a good partner for for us. We have been on the trajectory trajectory, uh, with them, um, for for a longer time now.”
Boeckem expects 2026 and 2027 to be the years when the focus shifts to deploying robots in real manufacturing scenarios.
“twenty six and twenty seven is probably the years where we, uh, maybe step back a bit and and and and and look for, um, deploying robots in actual manufacturing scenarios”
Boeckem poses a test in which a humanoid must find about twenty wrong gears in a box of gears, which he says is yet to be proven.
“I have a box full of gearbox, uh, o of of of gears, for example, and there there are maybe maybe twenty in there, um, that don't belong in the box.”
Boeckem says edge AI, meaning compute inside the robot itself, is decisive for full autonomy and learning on the job.
“However, to to be fully autonomous, to to to, um, to fulfill task, um, and and and basically also learn on the job, I think the edge edge AI is is decisive.”
Boeckem says current machine guidelines cover a fair part of the requirements but are probably not sufficient for autonomous systems.
“Because what we what we have currently is is machine guidelines, which probably are not that sufficient for the for the autonomous systems, although they they already cover fair part.”
Boeckem says many humanoids shown in demos are still teleoperated, meaning remotely controlled by people.
“still of the humanoids in in demos and so forth are, um, are teleoperated in in in in in these kind of things.”
Boeckem says robot use cases tested in labs are now moving onto true manufacturing sites, though not yet in dozens of robots.
“certain use cases were tested, uh, in a lab environment, and um, I would, I would, I would see currently that that this goes out now in onto a true manufacturing site”
Hexagon has a division called manufacturing intelligence that is working toward lights-out manufacturing, where factories run without humans.
“We have a division that calls is called manufacturing intelligence, and they're working towards what is called, uh, lights of manufacturing”
Hexagon developed scanners for robot dogs, which sometimes failed at tasks such as pushing an elevator button.
“and we developed scanners that went on on on on robot dogs. It sometimes, um, did not work because there was an elevator button to to push,”
Boeckem cites a digital twin of Notre-Dame de Paris as a famous example that supported the cathedral's reconstruction.
“basically the digital twin that allowed the the the reconstruction of not to Notre Dame in in in in Paris.”
Boeckem says that without a reliable digital twin of the environment, a company is probably funding a perpetual pilot.
“If you do not have a re- reliable picture of the reality you're in, you're probably funding a perpetual pilot.”
Boeckem says that even if the technology existed, safety certification for general-purpose robots in chaotic environments would be very hard.
“even if the technology would be there, it it would pr probably be hard to to to get the safety level and the safety certification right in in these environments.”
Boeckem defines physical AI as AI that acts in the real world and explains why it must be anchored in verified, accurate data.
“Physical AI is actually AI that shows up in the real world.”
Robots can train and fail safely in an accurate digital copy of the real world. The copy must be kept up to date.
“actually train the robot in this digital twin and and try things out”
Boeckem contrasts viral robot videos with the unglamorous work that creates value, such as inspection and material handling.
“It's not about showing the next coolest trick.”
Safe robots should recognize when they are lost and stop, and Boeckem names functional safety as 2026's big theme.
“the robot does know when it does not know anymore”
Boeckem says both cloud and edge compute are needed, with edge AI decisive for autonomy. He expects physical-AI regulation to grow.
“I would say both.”
Demos are often teleoperated, but constrained-environment autonomy looks close, while general-purpose robots are still far off.
“one, two, max three years away for autonomy in certain constrained areas”
Boeckem argues the main obstacles are fleet maintenance and enterprise readiness, not core technology.
“I would see enterprises not yet ready”
Humanoids suit workplaces built for humans, though fully autonomous lights-out factories would not need them.
“there's still a world that is made for humans.”
Boeckem warns boards to see robotics as a complex engineering system rather than a software rollout.
“boards overestimate the speed and underestimate the system work”
Constrained autonomy will scale where environments are mapped, while the general-purpose butler humanoid remains distant.
“the so-called butler humanoid is probably still”
An AI chatbot that hallucinates is annoying, but robots that screw up can cause death and injury. Burkhard Boeckem is CTO of Hexagon, which builds physical AI.
Physical AI is actually AI that shows up in the real world. It is embedded in machines that move, operate and affect people, assets and operations. Physical AI lets autonomous systems like drones or robot perceive, understand, reason, and performs actions in the real world.
The core aspect here is the operating in the real world. And when we talk about the AI aspect, where does that get overlaid into this?
It's important for physical AI and and specifically you just mentioned if a chatbot is hallucinating it it it's annoying, but if if a humanoid for for example, is hallucinating or getting out of control, it's, uh, it's, it's. It can be dangerous. So basically, the bar is much higher for physical AI because it it affects the real world. Um, that means, um, you have to ground your AI in in reality. That means, um, you use calibrated sensors, verified maps, do accurate reality capture in order to, uh, create a exact replica of the real world into digital twin.
And you constantly have to update this, so it's it's evergreen, and the system needs to be anchored in the truth. So that a robot knows, perceives, uh, things right, and can can, um, can handle things accordingly, or or perform action and task accordingly. But into the real world, and therefore you need this strong connection into the reality, or the strong link of the digital twin to the reality.
You've used the term digital twin a couple of times. How does the digital twin aspect relate to physical AI?
I would say, is the base, um, because robots need to to to learn, and they do it, uh, in various, uh, techniques and technologies, imitation learning, reinforcement learning and and so forth. But I, I would say the um, the, the, the s most straightforward and the um, the ones that is is is laying the bar highest for, um, function. Safety, for for instance, is is if you, if you take the reality, the real world, and and map it one to one into a digitality with a with a precision and accuracy that that guarantees you the ground truth that is is needed for the for the actions.
And then, ideally, your robot has on bo- on board a lot of uh sensors that uh f first of all perceive the world, and then basically can, um, can also control the manipulating or the handling, or the dexterity of the robot, or the locomotion, uh, where the where the robot is is going.
You have used the term ground truth. What do you mean by that? We're really covering the the basics of physical AI. And then, of course, we'll get more into the details. But this is super helpful to give us an understanding of a a CTO of Hexagon, which develops these products, uh, to give us a a basic context.
what is important, and also, uh, uh, where you base afterwards your simulations on and is is to have an understanding of the real world, of the of the real environment. And you can do beautiful things in in in terms that if you, if you accurately, um, create a digital twin by reality capture technologies, by vision technologies, for for example, and this is dimensionally accurate, this is accurate in scale and so forth. And then you can can actually train the robot in this digital twin and and try things out, make errors and and failures in the in the digital twin.
And then when you deploy it into the real world where you trained it on the digital twin, it's it's. It's far more easy, um, uh, to to to do this. The robot directly, um, uh, knows the environment, perceives the right things. And what I also find it's it, it, it has to have a certain sensor stack on board. Because, uh, when you created the digital twin, and we say for autonomous system, the digital twin should be autonomously updated, but that only works to a certain extent.
So for the for the things that changed, for example, a table or a chair that that have been moved, or basically a workplace that looks slightly different. Uh, at the time when you generated the digital twin, uh, the robot needs to have enough perception sensors and reality capture sensors on board in order to perceive the difference and and act accordingly. But I think the the dimensional ground truth, and I think that's that's where, uh, the interaction between the matter, the workspaces, uh, things you do, and also collaborating with with with actual humans, uh, start is is the ability to to base it all off of of of reality.
Um, and I think that's also a hard problem, um, that um, we need to solve, uh, uh. Furthermore, in the future is to close this gap between the the reality, the digital twin, and then the digital twin is is the sandbox where you simulate and train on in order to to close this real to sim gap again when deploying into into reality, because robots, in in essence, then live in reality in the real world.
Folks, if you're watching on LinkedIn, pop your questions into the LinkedIn chat. If you're watching on Twitter X, use the hashtag CXO talk. Burkhard, can you talk about some of the differences between AI, uh, digital AI on a computer versus what happens when AI operates in the physical world in areas like safety, cost, reliability?
It's also two very different, uh, different things, um, because, um, we all know the the the digital world, and we we know the the agents and the the chatbots, and uh, and and the co-pilots and so forth. But let's focus for for a second on the on the ph- uh, on the on the physical world, um, where robots and physical AI make it a difference or need to make a difference. I is is basically if they solve and um, and have you in humans to, um, to collaborate and and do meaningful things.
And and here we we talking about a a very different, uh, metrics, uh, for for, for example, um, we we are talking about, uh, predictability, we talk about uptime, we talk about, uh, safety, and um, that's something that uh, needs to be, uh, reliable, um, modeled into into the AI. And um, a a as I as I said before, um, we we're talking about, uh, maintain and security, uh, workflows, uh, that that has to, uh, have to be fulfilled, uh, in in order, so so it's, it's.
It's not about showing the next coolest trick. Also, I like this very much. If I, if I, if I see a a a robot doing doing something stunning, or have a very liquid motion and these kind of things. No, it's, it's, it's, it's probably first or first and foremost, uh, about, uh, the uptime, the the reliability and uh, and performing the workflows. Uh, we we intended it to to to do.
So it's really all about the the business applications and being very clearly focused on on the business problems that are trying to be solved.
Absolutely. You see, of course, a lot of YouTube videos where where, uh, humanoids or robots are launched that do super cool things, and it's it's so liquid these days. But I think the the really hard tasks, um, uh, look actually a little little bit more boring to to to be honest, is is is working in, for example, material handling, checking things, quality inspection. What, what, for example, we have shown when we launched, uh, uh, our our humanoid was, um, uh, our humanoid scanning, um, uh, with a with a laser scanner, uh, the um, the the door of a of a car for for defects.
Uh, it. It's probably far less cool and cool to watch, like a robot is dancing or d doing somersaults and these kind of things. But but these are the things that, um, um, we expect, um, specifically, humanoids to to do in the future, to, uh, to, to solve the tasks that are tedious, maybe maybe dangerous, uh, maybe in scenarios where you have a a high degree of, uh, repeatability. But the workplace is changing, uh, and and and these kind of things, and and probably, um, less about general purpose.
It's, it's, it's doing jobs, but being very flexible in in doing jobs. Otherwise, you could, um, uh, install a stationary robot for for just doing, uh, one of few things, and probably we will outperform, um, what we currently see in in humanoids.
Anthony Scriffignano, who is a data scientist and a prominent data scientist, and he has been a guest on CXO talk a number of times. He actually asks about physical a exploits by malefactors, and you mentioned security earlier, so can you talk a little bit about that aspect?
It starts already with the design of the of the of the whole robot and and is also basically in in how you guarantee functional safety along the way in all the aspects. In terms of um, it, it starts. Also, I I lately looked at a at a report where we have shown why why are the the current humanoids not giants? And and and these kind of things, um, so so it starts in building in things and building, uh, separating also also the the the control loop from the um, uh, from the uh, um, a AI, the physical AI parts of of of it.
But you have to be, uh, in a in a position, and and probably the best design goes in a in a direction is that that the robot does know when it does not know anymore, when when it's lost and slowing down, stopping, um, instead of of doing something, going on, tripling over, uh, uh, these kind of things. And also it's important by um, by design, that that that you, uh, design design, quite failsafe, you have that you have independency, redundancy in in what you you do, and um, and also have a have a sensor stack, um, that allows you to to perceive if things, um, go basically out of the range where you, where you trained it for.
Whether it's be by reinforcement or imitation learning, or other, uh, or other techniques in in in physical AI. But but I think, um, what is, what is important is that you control all the stack. It comes from the locomotion, the dexterity, um, the s the sensor stack, and then basically the physical AI, um, that that this is really failsafe, because, uh, it could could end up quite badly in a in a disaster. If if if you triple over on some, uh, on on on something, and, uh, and and people people might get hurt, and that is basically to avoid it at all cost.
In terms of of the I would say the big theme that is coming in 2026, which is functional safety.
Unlike digital AI, where in most instances an exploit can have an impact on a a network or have, uh, financial consequences. With physical AI, uh, there can be physical damage that's caused, so you have this additional layer of complexity.
Yes, absolutely. Manufacturers or or development teams of of of specifically humanoids or quadriplats or or drones, or any autonomous system that that interacts and is in in inherently dangerous in in in this regard. Um, you have to to to be able to to to control the whole system, not just the model. So it's about the sensors, the compute, the controls, integration, the maintenance. Even, um, it's nice if it's working out of the box, but but what is if if there's wear and tear, and over time, systems, uh, maybe not perform to the standards.
You you had them at the beginning, and then security, so all the workloads that, um, these workloads determine whether it it actually works in the in the in the real world. So I would say the bar is far more higher than than in in, as you call it, digital AI.
Greg Walters asks the question that I think lots of us have, which is at CES Computer Electronics Show. There have been lots of AI announcements from Atlas and Gemini to, uh, Nvidia. Do you have any comments and thoughts about all the announcements that took place at CES?
It's hard to comment on on on everyone, uh, on its on its own, because there there are so many. What, what, what I, what, what we currently currently see, and um, when we, when we started, for for for instance, our our humanoid project, and it was basically based off of our experience. We made with, uh, um, autonomous drones, and um, and what we, what we made of, uh, quadriplecs robots, or the robot dogs, as we was, we like to to call them that probably.
And and these announcements, uh, show the way, um, and probably with a with a never increasing pace, um, that when we, when we, when we started this, we, we, we thought, how in the world are we? Are we solving all this control loop? Um, tasks that that we have to do, um, and and with the with, with all the announcement, all the technology that that came, um, it, it brought more and more tools, uh, to to ba- basically build build robots.
And I would say, currently we are, we are. We don't see the end there, and I'm I'm, uh, what, what we we saw for Nvidia, the the the physical AI, uh, announcement around physical AI, and um, as Nvidia is a good partner for for us. We have been on the trajectory trajectory, uh, with them, um, for for a longer time now. But but it's, it's, it's, it's amazing to see, um, um, what is the capability, uh, today of physical AI and and and humanoids?
Yet I would say, uh, twenty six and twenty seven is probably the years where we, uh, maybe step back a bit and and look for, um, deploying robots in actual manufacturing scenarios, in in s in in workspace scenarios, in order to, um, to to solve real world, uh, problems. Because we have seen now that that locomotion is great, the liquidity we see in the motion, the fluidity, um, of of of robots is is stunning.
Uh, it's so stunning that you probably need to take them apart on the stage that you see that that there's not a a human in a in a in a costume, but yet to be seen. Uh, our task, um, I have a box full of gearbox, uh, o of of of gears, for example, and there there are maybe maybe twenty in there, um, that don't belong in the box. Can a can a humanoid sort them out by by looking at them?
Can they measure it? Can they see whether it's intolerance or yes, can it can be taken further? I I think that's that's yet to be proven. Um, uh, these these use cases, um, that these use cases can be fulfilled. I'm an absolute optimist in terms of robotics and and humanoids, and I think it's absolutely possible. But I I think I think this year and the next year is probably the ones where we where we see a lot of robots, humanoids specifically deployed in these real world scenarios.
Maybe maybe maybe still, um, isolated from the actual workforce, um, and then then doing this task, and um, you probably need then, uh, a more refined physical AI to to do specific things. Um, uh, starting with with with various, um, various interactions, already sorting and picking things, um, is is one that that we we should see. That is, that is quite doable in the in the in the future.
Let's jump to another question. Uh, this time back to Twitter to X. You can see I love taking the audience questions like you guys in the audience are so smart. And by the way, now would be an excellent time to subscribe to the CXO talk newsletter. Go to CXO talk dot com right now, subscribe to our newsletter and we'll keep you updated. Cause we have really incredible shows all the time coming up and you should join us.
Okay, here is a question from a technical question from Chris Peterson on Twitter, who says, Do you see embodied AI and robotic systems using more centralized cloud distributed edge or both? So is it cloud edge computing or both?
I would say both. The cloud part is is is, as we said in the very beginning of the of the discussion, is is very important when we when we talk very big data. In order to, um, to create digital twin, and then, um, the sandboxes for the the the robots to to to learn, and then also, uh, in order to orchestrate, um, maybe in the future. Currently, we have, uh, uh, a limited amount of of of robots, but when there's a large amount of robot, you need also to to orchestrate, uh, the the fleets, um, and that's probably you you all to do within a digital twin, um, environment.
However, to to be fully autonomous, to to to, um, to fulfill task, um, and and and basically also learn on the job, I think the edge edge AI is is decisive. That's why I think it's it's also a success story to to have sufficient compute, uh, in in your, in your, in your robot, um, but I think, um, together with something probably that that that is is in the direction of federated learning, federated AI, where one robot learns everything for the fleet.
Um, you, you probably need to to find a good balance between the cloud and the edge. Um, uh, uh, when, when, when there's then a deployment of a of a whole fleet. But I I think in in humanoids specifically, and or or or in robots in general, I think edge AI is is is really decisive, uh, for them, because, uh, the more compute, the the more things you you do. It's it's going more into the direction of a of a of a multipurpose, uh, robot, um, which I think is still a little bit out, but but I I think edge AI is is really important.
Have to have enough compute on the edge.
This is from Jotiram La Lad, who says most responsible AI regulations are written with digital AI in mind. Do you think today's regulatory frameworks are truly fit for physical AI, where AI decisions directly affect the real, the physical world? And Anthony Scrifinano asks the same question, essentially, he says, Are there emerging regulations beyond general AI in the context of physical AI? So both of them want to know about regulations and responsible AI, and how is it different when it comes to physical AI?
What we see now is is, and I I recently, I I gave an interview about this, um, function safety and the upcoming uh, reg- regulations, and I I dig deeper into the the topic. Maybe maybe to to keep it, to keep it on the on on this level. What we what we currently see is is probably what we've seen with the with the UAVs and the drones a couple of years back. That while the technology is emerging and the capabilities and the skills are are are coming up, then also the regulatory body, um, will, will, um, will come, come further.
We have, uh, uh, we have probably a very strict body already in the EU, um, uh, in order that, but but probably for, um, for safety regulation. In terms of PH physical AI, we see probably in this year and and next year, um, more more regulatory bodies coming up, or or basically, um, uh, many, many, manifesting, and, uh, I think that that function safety is is is still something that, uh, currently everybody is happy that the locomotion looks good and these kind of things.
But needs to address that, and it's, it's, it, it needs to address if you have a a larger amount of, uh, for example, humanoids working together, collaborating with with humans, and as I said, so I in short, yes, I I, uh, I. In anticipate that that we we see a rise in in the regulatory bodies, uh, for specifically, uh, physical AI. Because what we what we have currently is is machine guidelines, which probably are not that sufficient for the for the autonomous systems, although they they already cover fair part.
And we have a a related question from Arsalan Khan on Twitter, who says, AI algorithms use data to make decisions, and non-physical AI bias can potentially, uh, be corrected by updating the algorithm and the data. How do you tackle bias in physical AI? And he's asking also whether we need regulations, uh, he's also interested in the regulis- regulation question, so the issue of bias in physical AI.
It starts already with the accurate, uh, representation of the of the of the real world in in a digital twin. If you, um, if you basically solely uh, base your deployments and your trainings on synthetic data, without dimensional accuracy and without the the right scale factors and so forth. Um, I think that this is where you introduce the bias. Um, if you, if you train in scenarios that are are based on on digital twins, which are dimensionally accurate, um, and and um, and uh, a, uh, uh, an exact replica of the of the physical world, I think you can limit, um, this, this bias.
And then it's, it's, it's also one thing, um, that, uh, the the robot is itself is capable of of sensing the bias in in sensing, um, whether, um, yeah, for example, whether, uh, the the perceived step height is, is, is xyz, uh, with enough sensing power on on board, and that that also influence the the the compute. Then, um, I I think foremost these these are the are are the are the things. And then basically, once you once you found a bias, you have to close the loop and and bring it back.
But that's something, to be honest, that we did in other industries and in other technologies, uh, since quite some time. Talking about, for example, the quality inspection of a of a of an of a car for for for example. There basically you you close the lo- the loop from from the CAD drawing to the machine part. You inspect, and if it's out of tolerance, you, you, you alter your machine tool. And I think maybe this is very, uh, a very mechanical, um, uh, explanation of this, but but the same, uh, needs to hold true here as well.
Let's go to another question. This is from Sai Panemurru, and Sai says, How close are we really to autonomous operations, and what is the one thing still preventing us from getting there? And Greg Walters asks a very simple, s uh, very uh, related question. He says, This is the next stage. Embodied AI is general. Once it has all the six senses of a human, how close are we to this? So basically, w- how how close to autonomy and to the this idea of robots as really you helper humanoids?
Very interesting topic and a good, a good questions in in this. And I think that's what we what we currently see is probably still, uh, still of the humanoids in in demos and so forth are, um, are teleoperated in in these kind of things. Um, however, I think I think probably we are in certain tasks, not that far away from from from autonomy, if we basically constrain the environment, um, uh, and to to to some extent, um, for for me, it depends a little bit, um, um, on on.
What is the capability of the sensor stack of the of of the robot? Um, that said, um, we, we, we have currently, um, also, uh, also a lot of of of imaging devices, a lot of time of flight cameras, um, microphones, a and so forth, um, in in the in. The sensor stack, the sensor stack is not yet, um, uh, fully, fully there, where where basically humans, uh, humans are, to my knowledge, in in what we, what we currently, um, uh, see, um, on the on the market or in the prototype stages, um.
But I would say strong vision systems, strong lidar systems, um, combined with, um, with some speci- specialty systems that that do, for example, touch sensitivity in the in the end effectors. Enable us already now to to to basically, um, autonomously perform certain task, um, for example, picking things and and, and, um, and and placing things. I would say there we are already quite advanced, probably in the in in the longer term. Um, because I I think the whole humanoid, uh uh, businesses was was quite fascinated with with locomotion and and basic dexterity.
Right now, I think we see in the future much more sensors being built in, and that that enables then, um, autonomy, um, probably drones are a little further advanced. What we, what we, what we've seen here, um, in in this, but um, also the interaction with meta is is, of course, less, uh, here. But but I would carefully estimate there that we are, yeah, maybe maybe one, two, max three years away for autonomy in certain constrained areas.
Um, I would be very careful in saying general purpose everywhere or robot in a household can do everything. I would be careful with this, but in in in constrained envi- environments, it's it's foreseeable. Yeah,
Can you talk about the timeline for robots to create actual economic value, not just demos? And, of course, that's relating to related to what you were just discussing.
there are various stages and phases where, where, where we, where we in, um, probably what we've seen in the in in the last couple of months, and maybe maybe year, is, uh, uh, uh, a lot of announcements with, uh, uh, uh, robot manufacturers, and I would say early adopter, uh, um, companies, whether it's it's automotive and and so forth, and and I think in this stage, uh, certain use cases were tested, uh, in a lab environment, and um, I would, I would, I would see currently that that this goes out now in onto a true manufacturing site, but not in the in the quantities of dozens of of robot.
Maybe maybe a small amount in a in a, in a, in a, in a constrained part of the of the factories. I think that that would we can expect fully this year, yeah, in in in this, and then, uh, a a more, a more rolled out, uh, um, deployment, probably probably, than than than the year after, um, but still to some extent constrained, and to some extent limited to, um, to, um, uh, uh, certain certain use cases.
Can you describe, uh, where robotics deployments today most often fail? Is it technology limitations, integrations with existing systems, organizational readiness? Where, where are the obstacles today?
Probably the technology is is is already quite advanced, uh, eh, in in in this, but i it's, it's. It does not have a track record yet in terms of, um, deploying hundreds and thousands. So, so basically all the uh, all the teething errors, uh, the the teething problems that you that you would get with a with a big bigger fleet and all the maintenance and and all the the serviceability topics, uh, have probably not addressed in, um, in in in something.
If you compare it for for examples to, um, uh, to to to drones, or even in higher scale, um, uh, cars, uh, at at this at this point point in time, um, there b- basically, um, we, we. We yet need to see, um, how, how, basically, um, how maintenance intensive such humanoids is. But probably the the the failure level is higher, not on the technology and the capability part, but but probably for, um, I would see enterprises not yet ready, uh, for for basically, um, welcoming a humanoid f- uh, fleet and welcoming it to to work, um, uh, alongside with the um, with the with the workflows.
We just talked about the regulations. I I think there we need to do a a couple of steps more in order to have this, um, uh, this correct, um, and then basically on the on the on the other side is, um, is for me the the question still open. Then, um, how does enterprises, uh, corporates, um, basically incorporate them, uh, the the humanoids alongside with with with workers in a in a in operators, uh, uh, in a in an actual, uh, uh, factory or manufacturing, discrete manufacturing environment, um, probably probably.
The latter part is is the more decisive going forward, because technology will evolve and we we will see it quite quickly.
And Greg Walters on LinkedIn asks a question directly relating to this, he said, Aren't humanoid robots in, uh, human-like form soon to be obsolete in the manufacturing realm as new production lines will be designed to be run without humans?
We have a division that calls is called manufacturing intelligence, and they're working towards what is called, uh, lights of manufacturing, where basically manufacturing is, or the manufacturing environment is designed for having no humans at all in the in the loop. Um, and and even even even taking the lights off in order to, um, it runs autonomously, the the whole, um, the whole manufacturer. I I think for these environments, um, you, you probably would not, uh, develop a humanoid for, however, there's still a world that is made for humans.
There's still a large legacy, um, all over all of the world where basically sometimes, um, uh, whole shifts break down because a machine tool just to have to to be, uh, reseted, um, and uh, and these environments are are made for for for the last decades for for humans. So I think the proportion, the height and everything, eh, is is still very, very beneficial for human. But I give you 100% right. If you, if you design a fully autonomous lights of manufacturing plant, then probably humanoid would be not the desired form factor.
But for everything else, and there's a large legacy still existing, I would say the humanoid is is is a very, very nice form factor for this. Be because we have seen it in the there was an early hype for for these robot dogs, um, and and we developed scanners that went on on robot dogs. It sometimes, um, did not work because there was an elevator button to to push, or it was, um, um, uh, the not the right point of view, uh, for for the system, um, these are the the humanoids, uh, of course, over over overcoming, but I give you right for for a autonomously designed manufacturing place, you would probably work differently.
So the you the form, the humanoid form, could we say that it's essentially the user interface?
I would say, well, basically, uh, it's, uh, it's, uh, the um, the embodiment that resembles the user interface that we are used to in in this regard. Um, yeah,
So we have the robotic system that is independent of the particular form in which it's embodied, right? You could have a robotic system that's inside, uh, a drone, something flying, or a dog, or or a humanoid, or or, uh, a machine in a factory, essentially, right? So it's the packaging, it's the external packaging.
no, absolutely, and uh, uh, I I think what we what we still see is is that the um, uh, uh, in in in the humanoid there. They're not trying to to um, uh, to, to exaggerate certain qualities of a of a of a human, not making it it it much faster, not not longer legs, no slightly longer arms, these kind of things. Now we see a quite, quite an quite a nice, um, representation of a human as far as you wanna go, um, um, with, um.
W- without going into the uncanny valley situation.
Talk more about digital twins. You described earlier how the digital twin concept is foundational, but how are robotic and humanoid digital twins different or more complex from those used in industry? For example, in jet engine maintenance operations, which is an example that we hear about all the time.
What is is is very important for these autonomous systems and the and the and the digital twin. And and for example, certain a digital twin is always fit for purpose. For example, if you, if you want a digital twin of a jet engine, then then then certain qualities need need to be really, really in this digital twin. In order to drive a simulation in, in order to manufacture it, in order to lay out your production of the blades and blisk and whatever you you need for this.
What is important for for, um, for the digital twin, um, that you have, uh, in in the space of of robotics, is that, um, the autonomy of these systems requires that these digital twins have to be alive. Of course, they need to have the geometry, physics, um, uh, and and and the um, the right scale, however, um, you, you have to acknowledge the sensor behavior, um, the, uh, the dynamic obstacles, uh, that there are, um, and and basically they have to be constantly updated.
Because, uh, yeah, a digital twin, um, and yeah, famously, I think the the the probably the most, uh, famous digital twin years back was, uh, was basically the digital twin that allowed the the the reconstruction of not to Notre Dame in in Paris. Um, that was basically by accident or by luck. We had a digital twin, uh, of of this. This is an absolute static one, and you want to rebuild, um, uh, a building of this, uh, digital twins.
For for robots needs to be evergreen, always updated, and, uh, and sometimes even in in real time. And that's what what difference them, um, from, for example, the digital twins for for a part that you want to manufacture.
So, uh, Anthony Scrifinano is asking about, uh, collaboration, thinking, uh, among the, uh, robots, so he's thinking, for example, uh, drone swarms. Can you talk about that aspect of it?
That's something that is essential working in fleets, collaborating robots to robots, um, drones, drone system, like, like, um, in in swarms, um, as as I said, coming back to the to the, uh, for example, autonomous drones, uh, w- when there's a communication, or basically autonomous vehicle in mines, um, uh, yeah. The the communication between the robots, the autonomous systems is is essential, um, and it's essential for fulfilling the task. Sometimes it's also e- essential for, um, for learning together in terms of federated learning.
Um, if if one robot is is training something new, then basically you want the the whole fleet to to to have the skills as well.
Let's talk about some of the business implications here. What do boards of directors typically misunderstand about investments in robotics and physical AI?
But I think many, many boards overestimate the speed and underestimate the system work. Um, it's, it's, it's, it's not a a software rollout, it's it's a complex engineering system. That's what all also we we saw from when we talked about the regulation from digital AI to physical AI. It's safety engineering, it's operation change, it's integration and service and maintenance. And you have to think the whole, the whole cycle through, um, in in in order to to to execute, uh, on on a robot investment.
If you were advising a board considering a major robotics investment, what is the one thing you would insist that they understand before approving it?
I would insist on clarity about the job to be done and the risk envelope. Uh, I think you clearly need to know the task, the environment, safety constraints, and what happens when confidence drops. And I would also insist they treat data and digital twin readiness as a prerequisite. If you do not have a re- reliable picture of the reality you're in, you're probably funding a perpetual pilot.
Robotic efficiency or robot efficiency can mean job losses and displacements. Can you talk about the future of work in a increasingly robotic and humanoid robot world?
In an ideal sin scenario, where we where we currently work towards through, is the is the collaboration between humans and and robots. And we would like the humanoids to do things that are quite mundane, that are tedious, that are are dangerous, that are risky. Going into situations, whether it's mining or in or in the nuclear, where humans should not go. The so basically the the working together, and in in this scenario, I would say we we upskill the the humans to a to a different level.
And I just read an article about a famous robot assistant and said, I'm I'm optimistic about robots, but they should always work together with humans, and I believe in this as well.
What does responsible deployment of robots therefore look like?
For me, it means safety by design. You need the validation, the traceability, the failsafe, you need to have the the audits done, clear handover rules, and and and this is especially crucial in what I said before, in the especially in the mixed human robot environments. And it also means for me human centered integration, a training workflow redesign and the accountability to determine whether robots create value or or or maybe chaos in in here.
Arselan Khan asks the question, Do you see a time where robots become physical? AI becomes more and more autonomous, and will we need a manager, a a manager AI that checks if worker AIs are supposed to be doing what they are tasked to do?
Probably at some levels, yes, um, but but I would see that ultimately you should have somewhere a human in the loop. Uh, that is, that is, that is un unavoidable, whether it's then a management AI, or basically it's a it's a management AI, co-pilot that that helps a human to to to watch over a fleet. Um, yet to be seen, but uh, but but I think the the layering stacks of of AI probably probably makes a lot of sense.
Where do you see all of this going over the next two or three years?
That enterprises deploy more and more autonomy in in constrained domains. Inspection, intralogistics, picking, placing, handling, site monitoring and assisted manipulation will will scale in brackets when the environments are met, are mapped, or you have digital twins. And and the success matrix are are are really, really clear. What I would say, it's a little bit far away and probably still out of reach. Is general purpose solving everything, um, complete human level dexterity, uh, re reliable operation in in chaotic spaces and so forth.
Um, because probably, uh, even if the technology would be there, it it would pr probably be hard to to to get the safety level and the safety certification right in in these environments. So probably the the latter one is the is the hardest engineering problem. So I feel personally the so-called butler humanoid is probably still, uh, a way away.
Are there one or two core technical challenges that you are very much interested in and focused on right now?
If you have an accurate, uh, digital twin in order to to do the training right, in order to close the simulation to reality gap as as narrow as possible, that's, that's, that's one of them. Then basically the um, uh, the um, the the clarity on the use case and and the clarity on on what is out of scope. Uh, currently, it's, it's, it's, it's one other thing. And then then for me, it's, it's, it's, it's also.
What are the right sensors that I I put on a on a humanoid in order to fulfill all these tasks? And yet to have it as a as a cost level that is is creating a a economic value and a sufficient ROI in order to um, to to to see the invest um go through.
Well, this has been such a fascinating conversation. Burkhard Boeckem is CTO of Hexagon. Burkhard, thank you so much for being with us. It's been a really interesting conversation.
Thank you so much, Michael, for having me, thank you.
And everybody, thank you for watching. Now, before you go, now is the time to go to CXO talk dot com and subscribe to our newsletter so we can keep you up to date. And soon we're gonna be starting to send out summaries of our episodes on our newsletter, so check it out. We have awesome shows coming up and everybody, we'll see again next time. You guys, you guys are so smart. I always say that, cause it's really, really true.
Take care, everyone.