When Will Robots Be Better Surgeons Than Humans?

[00:00:00]In this long-term vision where this becomes a reality, not only can you train the robot to perform the surgery like the best surgeon in the world, you can also train this system to perform it like the best surgeon in the world on their best day. The eye is relatively rigid and if I perform this type of surgery, I can put suction cups and I can stabilize the eye really, really well. And once I have done this, I know precisely where I have to apply my plan.

[00:00:31]And so simply executing what I have pre-planned is going to be just fine.

[00:00:36]And there is not going to be a lot of difference. And so this type of of approach for delivering the care autonomously works really well. If you are thinking about building autonomy, the correct demonstrations are very important because of course you want to learn from the best surgeon how to perform a specific procedure.

[00:00:57]Welcome humans to the neuron AI explained. I'm Grant Harvey and today we are looking at one of the highest stakes applications of artificial intelligence autonomous surgery. Now before we get started and get into all of that, it's very exciting. Today's episode is sponsored by Dell Technologies Nvidia and you'll hear more about them in a little bit. So my guest today is Matias Umbarat, CTO and co-founder of Inner Logic and an associate professor of computer science at John Hopkins University. His work focuses on perception, simulation, and intelligence systems that could help surgical technology understand what is happening, determining what should happen next, and eventually take carefully controlled actions. Matias, welcome to the Neuron.

[00:01:39]>> Thank you so much for having me. It's a great pleasure to be here.

[00:01:43]>> Uh so, I guess let's just start with the phrase that probably makes some listeners uh excited and others deeply uncomfortable, which is autonomous surgery. uh when you use that term, what does it actually mean?

[00:01:56]>> That's a great question because I think it really um can have your fantasy run wild on what autonomous surgery might might really be meaning. I think in the very long-term future, I think you might truly be thinking about robotic systems performing certain subtasks or even procedures with with high levels of autonomy, fully autonomously um which is certainly a very long-term vision. But even earlier on, I think you you will see some level of these capabilities making their way into how we treat patients in surgical care today. In fact, you could argue that some of this technology is available on the market and is being actively used already today in surgery and we wouldn't necessarily think about those applications as as autonomous until we pause and you know reflect a little bit. But they are fundamentally enabling in some of the precision treatments that we offer today. If you think about um of themology for example in in LASIC where we try to you know shape the the the cora in order to correct for for vision for imperfect vision the way that this is being done with the laser and the ablation it's possible exclusively through automation because the precision that is required cannot be delivered by hand. Right? So some of these technologies are already part of of of the routine workflow when we think about this in soft tissue surgery particularly.

[00:03:35]It might still feel quite quite foreign.

[00:03:38]Uh but there are certainly reasons to believe that that this will be happening uh relatively soon.

[00:03:46]>> And um tell us a little bit more about your work, you know, at John Hopkins University and and how you're getting involved with autonomous surgery.

[00:03:55]>> Yeah. Um so at at Hopkins I think we've been at the very frontier of of of the science in the space of uh computer assisted surgery uh robot assisted surgery for for quite some time and this is not just me. We have uh a very strong center in in robotics that that specializes to a great degree in in medical robotics. Of course we have the hospital. So there's strong synergies and in there we've been you know like defining this frontier of surgical robotics and autonomy and surgery for for the last years and decades. Um it's not just me. There are many other colleagues that that are working in that space who are making incredible contributions uh to to this frontier. Of course they're also not just at Hopkins. there are many people, you know, in the US and worldwide that that are driving this advancements. But but for us, it's been really um demonstrating what we might be able to do um in in the near future and building out this vision and demonstrating that that this vision can become a reality. Of course, in in in science, it's more about defining what can be done. that doesn't necessarily mean that it's ready for productization. And this is what we're now doing in inner logic where we're taking our learnings that we had within our academic lab um and really trying to build loadbearing infrastructure that that helps us make surgical autonomy and precision surgery a part of that that is part of the of patient treatment in the future.

[00:05:40]>> Right. and um tell tell us a bit more about uh inner logic and and how that's how you've sort of branched off into that.

[00:05:50]>> Yeah. So um so at at Inner Inn in Inner Logic is is co-founded by uh three people who came together at at Hopkins is um Tito Poras who is the co-founder and CEO. He's a neurosurgery resident who uh did neurosurgery for a very long time all the way up to his eighth year in in residency. Um when he joined my my lab and was exposed to to the research that we're doing there at which point um I think he got more excited about uh the translational work that we could be doing together. Um and then two scientists uh Axel Kger who is a mechanical engineer and pioneer in autonomous surgery um and and myself uh both tenure professors at Hopkins who have been working together on on building complex systems for autonomous surgery in the past. And so uh we came together on on a quite interesting project around autonomous surgery at which point we decided that not only is there this opportunity to do trailblazing research but there's also this opportunity to really bring our learnings in in into this commercial ecosystem and make sure that other people can build on on our insights to to help us really move move the needle when it comes to providing better patient care and building new opportunities. for providing patient care across the ecosystem.

[00:07:15]>> So here's the thing about enterprise AI right now. Every leadership team has the pilot. Every company has the proof of concept. But actually scaling AI from the developer desktop to the data center all the way out to the edge. That's where things sometimes start to break.

[00:07:31]In fact, 95% of organizations say they can't even get their data ready for AI workloads. Not the model, just the data.

[00:07:39]And that's why we teamed up with Dellai Factory with Nvidia to build a resource hub we're actually excited about. It includes articles, videos, and case studies that cover the full AI journey from strategy and data foundations to infrastructure decisions to real world use cases and ROI. You'll also find stories on what an AI native factory looks like when physical AI hits the production floor to where hybrid AI workloads should actually run and what sovereign AI really means when you're operating under actual compliance constraints. And the whole thing is backed by the industry's first end-to-end enterprise AI portfolio. AI ready workstations, servers, storage, networking, all jointly engineered with NVIDIA. You can start small on a Promax workstation, scale out to the data center, and get to production up to 86% faster than going it alone. So, if you're tired of all that hype and just want to understand what it actually takes to make Genai work, head over to the enterprise guide to scalable AIHub on techrepublic.com or click the link in the description of this video. And now, back to our episode.

[00:08:43]>> Tell us a little bit more what what do you mean by um build builds on top of what you're uh what you're discovering.

[00:08:50]>> Yes.

[00:08:53]So in order to build out these autonomous capabilities of systems there is a whole lot of of engineering that that goes into this especially if you want to build these systems robustly. It goes all the way from understanding your databases, >> analyzing and processing the data in in a way that is useful and understandable then to humans and then uh to to machines and then translating that into how a autonomous system would perceive its environment, reason about what it perceives there and then ultimately perform action. Now this requires very sophisticated tooling at every single stage of of of the way and this is something that we have not seen build out and this is what uh at at Inn inner logic we're now trying to to help build which is computational development infrastructure for procedural medicine and procedural devices um including those that are in IA AI enabled um and fully autonomous and and and the Why I think here autonomy is the long-term goal but the mission is for procedural medicine in in general is because similar to what we have seen happen in in autonomous driving.

[00:10:18]I think the field will not witness a zero to one flip from there's no autonomy to oh today we're we're we're going to treat you autonomously. But it's it's it's a much more you know crawl walk run approach where we're we're starting with uh assistance systems and you know complimentary situational awareness that that is being provided by an algorithm instead of say a second person um translating into some assistive features that uh you know don't do the most safety critical and and outcome critical uh tasks but already offer some level of autonomy.

[00:10:58]and support ultimately to all all the way um autonomous subtasks. You could think about for example suturing um or tissue retraction and then more complicated as as as we advance. And so uh our goal here really is to to catalyze the ecosystem in in in building these types of functions on on the software on and on the AI level. um where our focus is not so much on building new devices because there are lots of of of different companies already in in in the space who are building phenomenal uh hardware and phenomenal robotic systems that are capable of treating new conditions.

[00:11:40]They're they're performing really well of uh offering new ways of of doing surgery. Um but because they are building a very sophisticated hardware product, equipping that with software requires a very different skill set and this is where we are going going to offer and help help them focus on on what they do best while also equipping their system with um a AI perception and autonomy capabilities.

[00:12:10]So at the same time that you are you know building this software you're kind of building it for I mean who knows how many different uh hardware projects. So you have to build it in a way that is almost like hardware neutral. Would you say that's accurate?

[00:12:28]>> That's absolutely right. I think one one thing that is very important in in uh medtech and here particularly also is is just differentiation because there are different like >> different providers care about different things the most right like there are certain ideas on how you think you will be able to deliver the best possible patient care. And so this is something that at the moment um every single provider thinks about slightly differently and this is certainly true but the the underlying problems are are fundamentally the same. How can I think about making my system robust to edge cases because there are certain anatomical varants that I just don't see often enough but I will need to be able to navigate and recognize reliably. So how this is a problem that that occurs doesn't matter what exactly the robot looks like the recognizing situations like that is important across the board and so exactly as you say for for us it is incredibly important to be uh flexible and customizable to the specific use context of a specific robot so that it's useful for for for that robot and and its manufacturer while also being sufficiently customizable that that we can adapt to those longtail edge events um that that you might might encounter in surgery in order to make these devices be it fully algorithms or all the way to autonomy uh robust in those cases where it truly matters.

[00:14:05]Do you think that there will ever be a industry push to some sort of standardization or you know like okay we figured out this is the ideal form factor for what these autonomous um surgery machines would look like or because you know you could customize it you know any number of ways um depending on what type of surgery you're doing there will always be this variety and you just have to accept that at face value and then build the best software possible that fits all these different use cases.

[00:14:35]Yeah, that's a good question. I think you you already see that there certainly is some form of gravitas to a certain kind of of of system form factor just because of certain constraints that the human body imposes on you. Um like like for example if um >> if you're thinking about laparoscopic surgery then what you will find is that essentially all robots whether they have one big tower on which there are multiple arms or whether they have multiple individual arms that are being positioned on the bed and then move towards the patient.

[00:15:14]There is v variability in all of these types of things, but at the end >> they all have one or multiple narrow instruments that have to somehow pierce ideally at a at at the smallest possible point um the skin >> and then have have to move and that that specific position where they pierce the skin that's the center of motion and that's where you know the instrument has to pivot around so that you don't harm the the tissue boundary right and so the design has to accommodate just this very natural constraint that if you're piercing the the skin somewhere, you really don't want to put too much strain and you don't want that hole to be big.

[00:15:55]And so as a consequence, right, like there are um constraints that that that simply imposes. So there is some I I would argue more natural selection on what what designs do make sense and which designs do do not necessarily make as much sense. On the other hand, I think it's it's also really exciting to see what type of new form factors, you know, the creative minds can can come up with because, you know, there are these ideas about swarms, about really nano robots that are act in very novel ways.

[00:16:27]And so there there's a lot of excitement uh uh to be had around where we might be going next. But if we're talking um you know conventional uh surgery where we where we interact on a macroscopic scale with um with with tissue the systems will look approximately just the same >> right. Yeah. Yeah. There's the the level of uh customization for your you know the creator's preference and also to make it differentiated enough but a certain point you got to have the same things to do surgery. Yeah. Makes sense.

[00:17:00]So most people are only using about 5% of what AI tools can actually do. They open chat GPT, ask a few questions, maybe write an email, and that's it.

[00:17:10]Meanwhile, other people are out there building workflows, automations, research systems, custom agents, and saving hours every single week. AI is moving so fast right now that most people don't even know which tools are worth learning anymore. ChatBT, Claude, Gemini, Codeex, AI agents, automations, five coding. It's a lot. And that's exactly why we built the Neuron Academy.

[00:17:32]We built a practical, self-paced AI learning platform that helps working professionals like you understand the tools that actually matter and how to use them in real work. Built by us, the team behind the Neuron newsletter and the podcast designed to help busy professionals actually level up their AI skills. Inside, you'll learn prompting, AI workflows, research systems, automations, productivity strategies, and practical use cases across the entire AI ecosystem. Everything is hands-on, beginnerfriendly, and built around realworld use cases you can apply immediately. These are real skills that help you work faster, think better, and stay ahead in AI. Right now, you can join the Neuron Academy for just $499 a year, unlocking 50 plus lessons from full AI courses to quick hit micro trainings you can apply immediately. So, if you're serious about leveling up your AI skills this year, this is where you need to start. Join the Neuron Academy today.

[00:18:28]>> I guess like to put this in perspective.

[00:18:30]So, where are we at today in terms of you know, you you gave us some really great examples in the beginning about how, you know, autonomous surgery is already, you know, parts of it are already being done. Um, what what would you say we're at in terms of how much is still surgeon controlled and how much is genuinely autonomous? Like are we like is it like a a 15 85%? Is it 7525?

[00:18:54]Is that not the right way to think about it? How how would you describe where we're at?

[00:18:59]>> Yes. So I I would argue that at this very moment we're probably close to 100% being search and controlled. And and the reason I say this is because even in in those types of procedures that I mentioned at the very beginning where >> the the way that the procedure is enabled is by having a system that can very precisely execute that essentially the this the specific um plan that this system is going to execute is predetermined by uh by by by a surgeon by by a provider who has made that plan and now the system is acting like a more sophisticated tool that that just performs instead of performing one incision in one specific cut, it just happens to perform, you know, very complicated interaction with the tissue.

[00:19:48]But it's all based on um it doesn't, you know, it doesn't act um or it doesn't plan autonomously if if if you will, right? Like it has very specific um boundaries within which it it it acts.

[00:20:04]And so I think that this is also why I'm saying like I think people would not think about a system like that necessarily as as autonomous because you have like probably you know what one might argue that if the outcome is deterministic because you already know what the system is doing. It is not so much autonomy it is automation.

[00:20:25]>> So you know exactly what is going to happen. Um and so in in in that sense I think what what you see right now pretty much all systems follow such a paradigm.

[00:20:36]So while they might move to a certain location fully autonomously there is no surprise so to speak in in what the system might be doing you know exactly right like this is what's going to happen um but these types of systems that have these capabilities we do see them used so I mentioned uh of themology is a great use case um there is radiotherapy for uh cancer treatment for example that has very similar profile where you know the the systems like Cyber Knife, the way that this is actuated is only possible through software and autonomous control, which is a fantastic accomplishment. There are um systems in orthopedic surgery where we can now perform precision cuts for for implants um that that have these types of of functionality. But but again right like it it depends on whether one perceives that to be autonomous or not because the plan still comes from a surgeon and and and then the the other systems that that one one has in mind especially for soft tissue surgery they they are operated in what we call you know tele manipulation where we have a patient side uh console that is robotic and we have a surgeon side console where you have the surgeon sit and and manipulate and One one can think about that essentially as a a remote controlled well device if if if you will where truly every single action that the robot executes on the patient side is input in exactly that same shape or form on on the surgeon side in in real time. And this is is fantastic because for example you can think about like long distance surgery which yeah which this for example makes possible but there again every action that the robot executes is in fact >> input by by a surgeon in real time. So uh while this is this form of surgery this robotic robotic surgery robot assisted surgery is becoming more prevalent in fact some procedures are are nowadays done close to exclusively robotically because of the benefits that this approach offers there is no autonomy in these types of procedures right it's it's it's fully telemipulated >> yeah that's amazing and you know I mean obviously the the kind of funny scenario you can imagine is like the doctor working from home now where he's like doing all the surgery from his house.

[00:23:08]But the real world application of that would be like imagine you know you can have access to the greatest surgeon in the world no matter where you live and you know if you can add autonomy on top of that we could potentially train all of the surgery robots around the world to do the surgery exactly how the greatest surgeon uh in the world does it and then in theory ever you know you could scale that to everyone. I don't know if that is plausible, but it I mean that would be like the dream scenario, right? Is that you could train it to even be better than the best surgeon, human surgeon in the world and and then you know scale that to whoever has access to the robots.

[00:23:45]>> Yeah. So you you're making a fantastic point and I think the the the point that I would add is you can in in this in this long-term vision where this becomes a reality. Not only can you train the robot to perform the surgery like the best surgeon in the world, you can also train this system to perform it like the best surgeon in the world on their best day repeatedly every >> right? And so this is where I think a lot of the opportunity comes in is because >> you're you're taking away some of those idiosyncrasies that just come with with the human nature, right? Sometimes even though you're very very good, you're you just don't have the best day. And so a slip up happens.

[00:24:33]A totally a robotic system does not necessarily have this type of failure mode if if developed correctly and if it offers the the associated robustness and and and therefore this long-term vision where indeed you have the capability of embodying the best possible human skill in something as scalable as an algorithm that can then be distributed everywhere across every hospital across every procedure that is being performed.

[00:25:03]in order to deliver to every patient the best possible care that at this moment in the world can be delivered I think is one of those big uh equitizers to to really health and and health access right and this is this is really fantastic um and ultimately why people are talking about autonomous surgery being a potentially incredibly exciting avenue because it it essentially by click of a button and once once developed you can deploy to every single robot that has the ability to execute which is really very nice.

[00:25:41]>> Yeah, because I was going to ask you earlier, I was saying, well, you know, it's it's actually pretty amazing that we can uh automate uh a deterministic uh program to perform surgery to our specifications. Right? So on some level I would ask myself, well then why why would we actually want it to be autonomous and why would we want it to uh be able to make its own its own plans? Um you know if we could if we could script it essentially the the way that we think it it needs to be done. Um but I think this is a perfect example of why we might want to do that. What other what other benefits does does autom actual autonomous surgery have that we're not thinking of?

[00:26:21]>> Yeah. So I think the the the biggest issue is not so much that we wouldn't want to script it that way if we could.

[00:26:30]The problem is that for most surgeries you can't. And and and the reason why that is is because the >> the overall tissue that you have to interact with is just not all that predictable. So if you think about >> interesting. Yeah. the the eye for example um the the eye is relatively rigid and if I perform this type of surgery I can put suction cups and I can stabilize the eye really really well and once I have done this I know precisely where I have to apply my plan and so simply executing what I have pre-planned like two hours earlier is going to be just fine because my eye is going to look like the eye looked two hours earlier and there is not going to be a lot of difference and so this type of of approach for delivering the care autonomously works really well. Another body part that has this beautiful property uh are bones. So if I'm interested in in in in putting, you know, like a personal and specific uh implant uh being the hip joint, for example, or the knee joint, >> well, I can take a scan whenever, even a week ago, I can make a very sophisticated plan of how I want to um you know, modify my my knee joint, how that implant has to connect, how I have to manipulate the bone during the surgery in order to put it perfectly with the right angle so that biomechanics are preserved for after the surgery and then during the surgery I just have to somehow align my robot with the anatomy and after that I can execute the plan and I'm done because the bone just doesn't change on you know an hourly basis right like this this uh abrasion on that which requires now surgery happened over years it didn't happen over overs and so for for many other procedures especially in soft tissue This unfortunately just isn't true because if you think about the the liver, the gallbladder, you know, the the abdomen in general, it it it's just not very rigid. If uh we we lie down on on the back versus on our belly, it will look completely different. So, if I go and I take a scan lying on my back and then I stand up and I wiggle around because I move to the bed where where now the that the surgery is supposed to be taking place, there's no guarantee that that it truly looks the exact same way how it did before. And so even though I >> we're too squishy.

[00:29:06]>> Exactly right. And and the problem is most of our most of our organs are squishy. Most of them are. And so for for these types of of procedures where you have to very delicately uh manipulate tissue, >> right? Think about um dissecting. I mean for for people who cook and and who interact with and who cook meat, there might be people who who have this feeling where they you know like trench and they like this is an interaction.

[00:29:35]It's very very complicated. It's very delicate. It requires very precise hand eye coordination. And as as you start interacting, you you don't really know what you're going to find next, right?

[00:29:45]Like, was this cut enough? Do I have to go deeper? Do I have to cut more? Do I have to cut less? And so, this is where you cannot make the plan ahead of time.

[00:29:54]You have an idea approximately what it is that you will be doing because if you know anatomy, you you you know what you will what you expect to find, but the specific approach, the specific plan, you still have to be able to adapt based on what you see. And this is why this this pre-planned approach doesn't work and why with this this advent of of of AI that we've been seeing over the last couple of years, we're now at a much better position to think about solving these types of problems because we we just have much better tools at our disposition in in order to think about right like how can we perceive how can we reason about what we perceive what it means about what we have to do and then control the robot. to to to actually execute it. Still very challenging, but compared to a couple of years ago, um we're now at a point where where where this is an a true possibility.

[00:30:52]>> I think of another scenario, and you could tell me how realistic this is, where let's say, you know, something unexpected happens. Maybe there's, you know, some radiation that like, you know, comes from the sun and like hits the chip of the robot at the exact wrong time and it flips a bit from zero to one and, you know, it makes a a incision it wasn't supposed to make. It would be nice if it was able to improvise in that moment and say, "Okay, I need to, you know, patch this um here or there or be able to or maybe there's a human in the in the room and they accidentally bump it while it's doing something uh you know, really dangerous and it makes that mistake and then it can course correct mid mid uh midsurgery."

[00:31:35]>> Well, it's it's it's an interesting point that you bring up. Uh I I haven't thought about it uh in in exactly the way that you describe it with uh you know like also the sun coming to to to allow us to play ball because there there are so many things that make this already incredibly hard. I I I I was hoping to keep the sun out of it, but the um but maybe I don't I I don't know.

[00:32:00]But but in fact, you know, I think the point that that you're making that that I think is an is an interesting observation is is the one that if you are thinking about building autonomy, the correct demonstrations are are very important because of course you want to learn from the best surgeon how to perform a specific procedure and this surgeon will have shown you because they perform surgeries and they deliver the best possible care day in day out. And so that you can learn.

[00:32:32]But >> right, >> autonomous systems and robots are not people. And the algorithms that drive them, even though we call it AI, and that gives you this idea that maybe they're similar to people because I try to train them on on demonstrations of real people, they just fail in very different ways, right? Like they fail in ways that people would probably never fail. they go in the in a little bit in the wrong direction and afterwards like if if you don't develop them correctly they they will get stuck because they now see a situation that they have never seen during training because the surgeon just never moved there right like the surgeon knows this is not where you go you go elsewhere but if if the system makes a mistake for various reasons because somebody bumped into it because the sun didn't play ball because the algorithm just didn't didn't get it right and it moved to a slightly location.

[00:33:28]>> How do you allow the system to recover from situations that in the real world you just don't necessarily observe? And so this I think is is is part of >> of of what we're building at inner logic also is the ability to synthesize these types of scenarios so that not only can you do can you observe what what goes right but you can observe what goes wrong, right? and and and you can do this at scale and and think about these edge cases, the recovery demonstrations and so on that will make these systems not not just performant but but resilient um under failure and under unexpected conditions.

[00:34:07]So this is kind of a a a funny example, but this is something I was thinking about, which is, you know, what if if these systems really if we get to the point, you know, this is a long path to get there, but if we get to the point where these systems really are able to provide that level of care that's like the best surgeon on the best day of their life every day um at scale, it begs the question, you know, what do the the human surgeons do, right? The people who've dedicated their career to this.

[00:34:33]One potential avenue for them is creating the the you know the training data of mistakes that that machines can learn on for forever where you know you have people work in in scenarios where they figure out what are all the failure cases of of what uh you know what how things could go wrong in surgery and then we continue to you know use that to feed the machines. But what what's your what's your take on you know what is the role what does the role of the human surgeon become once we get to this point?

[00:35:03]Yeah to totally I mean I think there is um there are certain cases so I I personally don't don't think that human surgeons will will will go anywhere and that's because I think there is you know the the way how I see and we we haven't talked about this so this is perhaps a good good moment to think about this is that already today >> we're we're unable to offer the the care that that really we probably should be able to offer right like I mean if you have uh uh if if somebody has a condition and they want to see a surgeon for for for some procedure to be done, >> the wait time is very very long, right?

[00:35:41]And so because of that weight time, people might not be following up. Um and then if you know if it's just something benign and it's really mo mostly for convenience that they were were to get this this procedure done, then maybe nothing bad happens. But there are people who are lost to followup simply because they cannot schedule the the care that in an ideal world they they need simply because the hospitals are you know like at at capacity and they cannot they they just cannot offer you know sufficiently like these appointments sufficiently soon. Um, and there are other cases where it's not just about can we schedule people. It's emergency procedures that people might not be able to receive because the hospital that is closest to where where the accident happened just didn't have the, you know, the right expert in house or the right expert there at the moment that this happened >> and and so they weren't able to receive life-saving care that uh would have gotten the much much better outcomes.

[00:36:40]stroke is such an example, right? where um not not every hospital h has an has the right specialty on on you know on on on staff and so then people are evaluated and you determine that they need throbectomy so reprofusion of of of the brain but it cannot be performed at the hospital where they are right now and then they have to be transported to a hospital where they can perform this type of care which delays treatment and therefore deteriorate outcomes. um usually because what they say in stroke is time is brain. A lot of people have heard that that phrase. And so if you think about the ability of having a a a robotic system there >> in in that stroke example, right, where the system has where that hospital has access to an endovvascular robot that that can perform the procedure regardless of whether they have a radiologist on or an interventional radiologist on on staff. And right there in the moment, the outcome might be very very different for for for those for those types of patients. And so I think that it is not just about are we replacing surgeons. I think it's the it's it's it's the wrong question to ask. It is about can we offer the the the care to the people who need it in in a in a sustainable version. And so what what I think will is is most likely to happen is that some of the more routine uh care that right now is is is being caught by you know patients by by by surgeons and is being performed that they might not find particularly challenging um that that some of those aspects of those surgeries maybe not all of it but some parts of it for example exposure uh where you make the initial cuts and you prepare everything for the safety critical parts of of of the surgery or uh the wound closure at the end, right? Like putting stitches and making sure that everything is fully sterile and and and and completely uh ready for for draping to finish the patient up. That some of those components can be performed with the systems help or by the system uh directly which essentially just frees up bandwidth in in order to treat more people in order to provide more people with that same care. And I think that there are developments that will require us to rethink how care is being delivered simply because we we do have an aging population which will require more care, which is fantastic, right?

[00:39:11]We're expanding the the lifespan and health span of of people, which is great, right? Like people live to older ages and they can do healthily, which is great. But part of that also means that the reason they >> they get there is because we're able to to treat conditions. And so we we need to continue to be able to do this especially as a larger proportion um is is older and might require more care.

[00:39:35]And then there is another thing that is fantastic in principle but challenging if if if presented in this context which is the fact that we develop new procedures and we can treat more conditions. So in fact not not just can we offer more of the same we can offer more of different procedures. And so just in general we can offer more >> right >> and how we offer more in in a you know surgeon constrained bandwidth and and you know capacity constrained environment.

[00:40:07]>> This is not exclusively an do we do this autonomously or do we not do it autonomously type question. But I think it is clear to me at least that advances in technology and advances in engineering these health systems and the way how we provide care differently is going to play a humongous role in in making sure that that we can offer the best possible care and we can expand what what is possible to be offered in in in the health systems. And so I think in that space I think technology and autonomy will will play a role. Not the only one. Not the only one. It would be >> naive to assume that that just introducing autonomy immediately solves.

[00:40:47]So it it will not, right? It's it's very healthcare is very complicated.

[00:40:51]>> But it it will play a role simply because we have a set amount of people. We only train I don't know how many surgeons, not enough. And and that then we make them work very very hard, right? Like we strain them really really badly. And so I think being able to to to make their life better so that they can do what they care about more which is giving patients the best possible outcomes I I think is is a great opportunity for us to advance health and patient outcomes.

[00:41:21]>> Totally. It also makes me think everything that you just said about radiology and I think that radiology is kind of considered the canary and the coal mine for AI automation and healthcare and supposedly you know the AI models that are able to uh or is it radiology? Um yes yes yes radiologist yeah um it's the can canary in the coal mine because you know all these models are getting so good at identifying you know cancer and and all of these other things that you're looking for um with the X-ray but actually you know there's more radiologists than there ever has been because of the technology you know the technology has gotten so good so more people can do it makes me wonder if a similar thing would happen with surgery where you know if we perhaps you know not necessarily lower the barrier for entry that we expect of our surgeons.

[00:42:08]But you know the surgery process you know is augmented with the this technology. Perhaps then more people go into surgery because there's more job availability because everywhere you know every hospital now can have these robots that can do x amount of the work and then you can actually have a surgeon on staff um you know in these places that maybe wouldn't have them before that can do you know a lot of the surgeries that you know perhaps only certain experts could do. Maybe we have more s we can support more surgeons um and you know actually solve that labor problem.

[00:42:40]>> Um abs absolutely I think because of healthcare and you know the world really but these processes I think they are so tightly interconnected that it is always very difficult to predict what turning you know one little knob a little bit what what the downstream effects might be. Um and so I think exactly as you in in radiology where >> it's radiology is is is nice because every single image comes with a report.

[00:43:11]So essentially you have the sample right this is this is what I see and and this is what the radiologist saw and the images are on grids so they're very very accessible for a uh computerized system to read and and and process and it happens to be a computer vision task where machine learning tended to be the strongest from the outset right back back from 2011 very very is really good um >> but I think a lot of things that have happened exactly as you say is that it increases demand because the first thing that you automate is not the the the fully autonomous read. It is workflow related and so now I can in fact offer considerably more reads to to to patients and so on. And so that that drives up the volume or I um enhance the the efficiency with which these scans can be taken because I have better reconstruction algorithms. I have a better way of and so now I just increase the volume and that suddenly increases the demand because I I can now offer uh similar care. So I think a lot of these effects might be happening and I would argue that in in surgery we will see very similar tasks or similar effects as well where I think that just because we will be able to to treat more patients the surgery adjacent professions whether it it's going to be the the surgeon or whether it's going to be a new type of of of job that that is related to this um is is perhaps going to to do exactly what you say is like grow and create more opportunities for for people to go into that field. But I think healthcare just generally I think is has these very complicated cases where where people this is the problem with illness and disease right is that the biggest issue about that is that it is it is a diversion from a pattern. So you know and and we call this machine learning previously we called it pattern recognition well AI and machine learning right and then previously when there wasn't not not as much hype people like to call it pattern recognition and ultimately not much has changed from from from from that pattern recognition uh uh way of solving the problem but exactly connecting this back is that disease and illness is a deviation of a pattern and and that makes it very hard to develop these types of algorithms robustly for for something that that is different than the norm because the algorithm by definition tries to learn the norm and and and not the deviation. And so there is like how how we offer the the care then also is is is going to be affected by what these types of systems will be able to do really really well and what they might not be able to do um as well as we are hoping uh them them to be able to do it. For example, decision-m under uncertainty might be something that we we just don't want to to to give to to a system. Um because we want to have the the human touch and the human decision- making and you know everything that is human ethics and values and things like that. We we we want to have that decision making process as part of it because as society we feel better about these decisions being made by by by humans. There there's research on on on these types of things. What what people tend to be okay with being decided by an algorithm versus where where they would want human judgment in in case and they tend to be forgiving even if the human is wrong because some decisions you just they're just hard and and so it's very difficult to automate them. So this is where I think it will be close to I I I don't see humans being taken out of of the care delivery s simply because when people seek care, they're vulnerable. The connection to people who care is is just as important as the actual procedure that is being delivered. And so I I I don't see that going anywhere. On on the other hand, as we discussed earlier, people have variability. Some days they you know, you you feel like you can only win, right? Like we have days like that and then there are other days where you feel like this is just not it's >> today it's it's it's not it, right? And the the problem is is is there when you perform surgery, um you not having the best possible day means somebody else not receiving the surgery that they could have received the next day. And so there having these assistive systems that just back you up and make sure that uh you having a bad day doesn't result in a patient having a terrible one. Um is is is really a first great way of of of just you know making sure that you're you're not taking out the the humans.

[00:48:15]You're just making sure that that you know they they they are they have complimentary awareness. They they they you you try to support them deliver the best possible care which they want.

[00:48:25]Yeah. Because again, right, surgeons and and clinicians in general ubiquitously, the one thing that they care about the most is making sure that they do the best for their patients. And so, you know, when you offer them the ability to do exactly that, they will not be technology adverse. they will will will not they will be critical as they should be because they're not playing a game.

[00:48:50]>> Yeah.

[00:48:50]>> They're they're trying to treat pe people. So they will have hard questions and you better have good answers. But once once you can provide the corresponding data basis and evidence that truly the incorporation of technology like that helps them treat their patients better. the clinicians will be the first one to advocate for for this type of technology to be used routinely because that's what they care about doing better for their patients.

[00:49:20]So this makes me wonder, you've you've described and I I really agree with and appreciate the vision that you've painted here as autonomy as like a ladder where you're slowly going up the rung, you know, one rung at a time and you're, you know, finding the systems where you can actually introduce that automation and where that does streamline things and then you slowly but slowly but surely, you know, go up the ladder and build up to, you know, potentially a fully autonomous surgery.

[00:49:47]um what is the first what are the first use cases that you're targeting at inner logic and you know do you plan to build any of your own you know AI models for that will you use other models and and yeah what's what's your first target goal here that you're trying to uh go after with inner logic >> so I think we're we're seeing the full spectrum of where autonomy at the moment is is the most mature and and most likely to to convert soonest for for some of these applications. And I think >> the the ones that that we've been seeing I think are the ones where robotic systems are already quite mature and ready to take the next step in in the level of autonomy that they that they provide. And then in in the area where not acting is worse than acting with a little bit of a mistake. So to make this a little bit less opaque, I think the first area is is muscularkeeletal types of robotics where we're dealing with um hard like rigid structures like bones where we we have much better predictability of how the bones will move. Um the instruments that are being used there also are are usually rigid and and and relatively bulky and so tracking them and and understanding where they are and making the right is is relatively easy because the perception problem for for the system is is is just relatively easy. So that that's one area muscularkeeletal applications and muscularkeeletal devices and procedures is is one of the areas. The second one that uh is important for for for us is is those im emergency type of of procedures where we're unlocking care in in those environments is is really important. So there we're focusing a lot on endovvascular type of applications because of this stroke use case. There is a a large interest also right now in endovvascular robotics that are being developed for these types of use cases. And so enabling that type of treatment is is important. Uh similarly to to disaster response where you can think about um conflict or other type of trauma that might happen where people are hemodynamically unstable. So, uh, bleeding and being able to treat conditions like that, uh, to stabilize patients and make them, uh, ready for for transport if if an expert surgeon to do that is is unavailable is is another one of those capabilities that are potentially quite enabling. So um these are I think some of the use cases that for the demonstrations where where we want to showcase some of the tooling that we're building that that we see u interest and that we see opportunities.

[00:52:30]But um again, we're we're building computational infrastructure that that is designed to help OEMs and and leaders in in in that space solve their problems that they have in procedural devices and procedural autonomy at at software speed. And so it is all designed to adapt to the use cases that um you know our OEM partners seek to solve using the platform that we offer.

[00:52:57]>> That's awesome. I'm very excited. I I I can't wait to see, you know, what how how you apply the the software tooling that you're building and and you know, how you can scale these systems because I I I definitely resonate with your vision. I think, you know, the more health health systems and, you know, individual hospitals and and doctor's offices where we can get this capability, I think it's all for the better because you can give that care and you can you can equalize it and get it out to as many people as possible. I think it's very important and better software helps us do that. Um for listeners who want to follow Inner Logic's work, where should they go and uh what should they expect to hear from the company next?

[00:53:37]>> Yeah, so I think we um we will be releasing some of uh our initial um technical reports and and other types of of of information about the company soon. I think we've been um not not quite operating in stealth until now but we also had haven't formally announced and so now there is uh the big uh conference for the society of robotic surgery happening uh soon where uh we will be uh present and uh you know presenting some of of our initial developments. So that's going to be exciting. So on on on the web uh and on LinkedIn is going to be the the best um outlet to to follow us. I will say on the idea of autonomy being a ladder and working our way up you know self-driv you brought up self-driving earlier which is a good example the the one issue with self-driving which I think is very has nice parallels to surgery is that you know driving a car is a very dangerous activity actually and there was a lot of examples you know of Tesla when Tesla was on the road and testing things originally where there was a lot of accidents and it got a lot of bad press But Whimo was sort of like a leapfrog product where it came out of the gate. And yes, of course, they had testing, but you know, comes out and it pretty much works at this point. But there's still times when even Whimo, which I think is probably the best self-driving product on the road today.

[00:55:00]You know, Whimos will still get stuck somewhere and you know, people you know, you know, take the meme of it and you know, they film it and everything. And I worry, you know, what what what uh you know, obviously when you hear autonomous surgery, you worry about that use case.

[00:55:16]And you really don't want your autonomous robot to get stuck midsurgery and then, you know, you have to have a mechanic on staff or, you know, somebody who understands the software to be able to actually like course correct it. What do you do in that sort of situation? Is that a reasonable parallel? And how are you thinking about it?

[00:55:33]Yeah, it's it's it's a great analogy and I think pe people have a very clear perspective on autonomous driving because it's been around for for so long and so we've experienced it in in in certain ways and um I think first of all I will say uh I I personally think Whimo is is a great product. I I I I use it a lot and even my kids are convinced that it's fun. Um but of course they the the the technology can fail. Um and and then it usually makes makes the rounds and and and people like to to to joke about it. What what I personally think however is that that reaction is is is perhaps not the best one because it's true right like if we deploy these types of systems at scale they sometimes just will get it wrong and they will make a mistake and then of course we can run our mouths and we can laugh about it. we can, you know, be be be um cynical, which um as as a German is close to my heart. I understand, but it it somehow misses the bigger picture. And the bigger picture, I think, is that when you look at the data, and there was this uh fantastic um op-ed in the New York Times late last year that was describing autonomous driving as one of the most important public health interventions in in the recent past because of how much safer it is compared to uh the the data of of real world drivers, right? And that's because yes, the system still fails, but hey, normal drivers, like human drivers, they fail a lot all the time also. And so I think this um this laser focus on on individual mistakes, I think, is just getting the big picture of the story wrong. And I'm I'm a little that the reason I want to bring this up is because as we are starting to introduce this type of technology in surgery, which admittedly is a very high-risisk domain and we need to make sure that everything is is is done up to standard and validated and done done safely, of course, but just because of how we cannot avoid mistakes, medical mistakes in regular practice because it's hard, we will very likely not be able to avoid individual mistakes sometimes either.

[00:57:51]And so this is where I see some of these parallels where we will have to do a really good job um in communicating and convincing that that this type of technology ultimately at at a population scale is is is going to offer uh tremendous benefits. Now again we're not there today and it will take us time both in terms of building the systems to that level of standard where they can per perform at that level and then demonstrating that truly they are safe which will not be tomorrow. It will take a while, but I think I there is going to be very likely a similar type of story where we will appreciate the introduction of these assistance systems and a autonomy subtask features as a similarly transformative public health intervention in order to ensure better patient outcomes.

[00:58:44]>> Do you ever think it will be seen as unethical to let a human perform a surgery versus a autonomous surgeon when we get to that point?

[00:58:55]This is not uh not not not an area that that I can speak with uh with authority.

[00:59:00]But I think that already today we see that there are certain tasks in surgery that that are very mechanical and and um like they feel repetitive. So for example, if if you have to perform different sutures for anesmosis where you join two tubular ends to together, we know that if you can automate this robotically, the rep repetitiveness and the reproducibility of the different spacing within between the stitches, the robot can just do it much better than the human because you know it is robotic. And so it it it does it with a much higher level of precision which creates a much more laminer flow profile, enhances the overall pressure that that can be applied. And so it's it's just better. It is there are proof points like this. And so it's it's quite possible that truly once this is validated as as safe and as as a real product level and ready for deployment that we we would feel silly not doing it the the new way because it's just demonstrated to be better than than what we can do otherwise.

[01:00:10]And maybe that maybe that that threshold to meet would be lower than like the self-driving threshold because not everyone is a surgeon, right? So, not everyone has a certain level of of expectation or self-confidence in their own ability like they do with driving.

[01:00:25]Whereas, I think driving, you know, has a much higher bar to cross for the regular average person because they're like, well, I drive and I think I'm a safe driver whether or not they are. So, they have a certain level of like, well, this has to be above and beyond safer or I before I accept it. Whereas surgery, perhaps the bar is maybe lower for the general person because they don't do surgery. So they don't have any sort of you know uh threshold with which has to be crossed before they'll trust it. I I guess obviously you need to know they can do it without without cutting you open and and all this stuff first but yeah >> I mean >> cutting the way they're not supposed to.

[01:01:03]I mean, >> yeah. I I mean, surgery is a is a highly regulated field and so I think once what, you know, once you really make it through the different regulatory bodies and and into a into a product, there there's a good body of evidence that um you know, the system is is is up to spec and if used correctly will will perform as indicated. Now, now this being said, I think there there is a lot of conversation to be had in order to to build trust about this type of technology because while the average person is not a surgeon and therefore might not have the same level of appreciation of what makes or doesn't make uh safety, I think people are also quite concerned about their health and they don't want like this concept of trust on what care they are being provided with I think is incredibly important and and whether we're we're we're ready to accept a machine performing parts of this. I think it will require a little bit of uh of conversation. However, if you look at elective procedures where you know people don't need the surgery now they can choose where they receive it. We already see right that patients vote with their feet and go to centers where robotic surgery for the procedures that they is being offered because of those demonstrated benefits of you know reduced tissue trauma, shorter recovery times um and so on. And so this is already an effect that that that we see. It happens.

[01:02:31]>> Wow. And doing that goes to >> that's that's awesome. I didn't know about that. So what so you know obviously I' I we we follow at the neuron pretty closely you know the developments of you know large language models and how they're progressing and now we're really judging not necessarily based on the model you know itself and how it responds but how it works inside an agentic system. Um and robots you know famously are a very difficult problem and you know a lot of people say you know in order to solve robotics you need to solve robotics. Um what would you say is the hardest part about um you know engineering uh for for you know robots especially surgery robots in this case um and you know how are you trying to approach that and and do it differently? I I think by by far the the biggest issue is that you have right right now you have robotic systems that are designed the way that they are designed. We have this big benefit in surgery. We have this big benefit that in fact the data flywheel exists. So we already have surgeons that are using these robots to perform surgery every day. And so this demonstration data uh exists. That's that's one of the big benefits that that we have in surgery compared to say general purpose robotics where you know data on how to fold a t-shirt using a robotic system it just doesn't exist. And so if I want to automate that with a humanoid robot my first step is to go and acquire I don't know how many demonstrations of of folding t-shirts which is not a particularly pleasant thing to do. Um but it's important because that's what what what unlocks the task. Now in surgery again we we don't have this problem because we already have surgeons acquiring this type of data. But the the visibility of structures, the knowledge required in order to do surgery safely, the delicacy with which the tissue needs to be manipulated precisely. the viewpoints that have to be selected in order to do it safely. I think all of these are are incredibly hard challenges that while we're seeing that in a very narrow context we can automate them. So we've seen clipping and and and cutting for example of the bioduct fully autonomously um in in live animals. Um, so while we're we're getting closer, doing this robustly, especially across potentially diseased and therefore very inaccessible anatomy is incredibly hard.

[01:05:13]And so this is where I think like we will first see some of those more predictable type of of procedures that that are being automated and those assistive tasks that don't even require necessarily um tissue manipulation like camera control and things things like this that they will happen first. But yeah, I mean uh exactly as you say uh in in order to solve surgical robotics, you you you have to solve surgical robotics and that that that is about as hard. I think with what we're building at inner logic our hope is to take a lot of it from realworld experimentation which by design is slow because you have to build a mature prototype. you have to do things like it you can acquire with a real physical prototype you acquire data at at 60 minutes an hour right like it doesn't it doesn't go faster unless you build a prototype >> now we're trying to take a lot of this >> right >> real data the real data will not go anywhere because again if you want to learn what the best surgeon does you got to learn from what that best surgeon did this real data but for a lot of it like the demonstrations about recovery padding it with with having the system understand robustness and and and so on a lot better. A lot of that can be taken in into software and that allows us to speed up some of this development and especially speed up iteration cycles.

[01:06:36]And so this is one part of what we're building at in inner logic in order to make the development of of of these types of of algorithms and systems that power this this push towards uh autonomy and better better procedural devices to make it less painful and allow us to iterate faster.

[01:06:59]Is there a role for uh like sensor data in that? Like in the same way that Whimo, you know, uses a lot of different um I forget exactly what they're called, but they use a lot of different sensors beyond just cameras. Is there, you know, do any of the robots that exist today?

[01:07:13]Do they collect multiple different types of sensor data when they're doing these surgeries? Could that be useful to the process? I know that's a lot more data than you have to process on the software side, but um >> No, no, absolutely. In fact, one one of the big challenges when surgeons went from conventional laparoscopic surgery where you take these long uh scissor-l like instruments with long shafts and and you manipulate like this and you know for minimally invasive surgery.

[01:07:42]When they went from from those types of of of procedures to robotic, one of the biggest issues was that one of the sensors that your hands had built in suddenly was gone. And that was force feedback because the robot does not communicate to you how how strong of a force you you are applying. And so you have to infer the the strain that is being applied to the tissue from the visual from the visual system, right? So you have that visual haptics how how hard are you pulling based on the deformation that you observe uh with your eyes. That is something that has to be learned. And so we we train surgeons to do it this way. But no, new systems now bring for example these four sensors back in into the device and so you know quite precisely how strong you're pulling um and you can incorporate that into decision- making but as as I said earlier right like >> right >> this differentiation that individual manufacturers have for their devices and how they want to use it is is different and so we're we're accommodating it all depending on uh the specific use case and and and partner that >> cool very cool Well, I wish you all the luck with that because I think, you know, general a general uh surgery robot would system would be really amazing and could benefit everyone, you know, especially because you're if you're learning from every system and learn how to work with every system, I think that would be the best case scenario because then you're building the best system possible that everyone can benefit from.

[01:09:11]So, really appreciate what you're doing.

[01:09:13]Really appreciate you taking the time and uh Matias, thank you for joining us today.

[01:09:18]>> Yeah, thank you so much. This was fantastic. Real quick, I want to take a second to uh offer our thanks again to Dell Technologies and Nvidia for sponsoring today's episode. For our full content hub on Aative Factories, hybrid AI workloads, data readiness, and sovereign AI, head over to the enterprise guide to scalable AI hub on techrepublic.com or click the link in the description of this video for more.

[01:09:39]If you like today's video, please take a moment to like and subscribe. And don't forget to check out the Neurons other projects, including our daily newsletter, which is read by more than 700,000 people just like you, the Neuron Academy, where you can learn all about AI and how to use it in your work and life. And also check out our new sister newsletter if this topic is interesting to you, Robotics Insider. Thanks for joining us, and we hope to have you back next time. Farewell for now, humans.