From Autonomous Robots to Physical AI

Thanks a lot for the invitation, and I'm happy to join here. And as Reggie introduced already, I'll focus more on the theoretical part, but I'll have practical bits as well.

Today, if somebody would come and say we have the most talented team and we developed this surgical robot that works fully autonomously, would you agree to be the patient? If they would let me see how it performs for a couple of surgeries, I would be more than happy.

me you know so that's basically what i do is to try to get people working in technology in this case miral who is actually kind of the mastermind behind the table tennis robot that can beat elite players that was developed by sony ai and to share the know -how how they actually did it and just

Just this Monday, there was a new list released, which I had no idea about before last week, actually, of top 100 women in AI, data, and robotics. And she was on the cover.

And it's amazing to actually have somebody to join the interview and to explain what they do. And then suddenly, they become a bit of a superstar.

And nowadays, we see that AI, which lived inside the computer optimizing the processes, is actually leaving the screen in a way that a lot of things are shifting towards physical AI.

And instead of giving my opinion on that, what I'll try to do is to summarize the first 10 guests that I had on my podcast, which is called Tech Drive Zurich, and show their insights that they gave in developing AI for the physical world.

TechDrive Zurich: Conversations Beyond the Stage

And just a few words about myself. So my name is Justyna Meshakes and I have actually a PhD in robotics and computer vision and EMBA degree and as my full -time day job what I do is basically I try to turn AI research into the products and this show

was motivated by attending many conferences where you have a really nice and impressive talks about what people do how they solve problems but I found very interesting is where basically the talks after the conferences when you go for dinner you grab a beer or you actually catch a train or a ride back home what they tell of behind the scenes and that's exactly what I tried to capture not what they build but how and why they

do it and this is a very short teaser still work in progress how the show looks like a little bit

Using AI to Produce the Show

so just very short new intro on this and as there was one point about kind of the questions and these meetups of how ai is being used in personal projects so this is i would say semi -personal project i decided to give a quick glimpse into how i use ai to actually create the show

because as i mentioned i do have a i have a full -time job then i have two small kids and and also a couple of other side hustles. So there is no chance that you could do from filming to the final episodes as one man show before, like three, four years ago, without the small team behind you.

But now it's possible to pull it off by basically combining a bunch of these tools. So Cloud Code and ChatGPT Codex are now my best friends in production.

The Production Workflow

And just to very briefly explain how the process looks, so I have a couple of cameras in the car a couple of microphones each for one speaker you have to sync use bunch of libraries to do that you do the transcription because the meaning is really inside the text and as you can

imagine driving in the car around the city sometimes even hitting random gravel road or a tunnel or something like this then you have bunch of random noises and you need to remove the road noise as well then cutting you want to to focus on the person who is speaking,

and not just the same scene all the time, because we have short attention span, so we get bored.

And the cool part is that I started using DaVinci Resolve Studio. And just a couple of weeks ago, in the paid version, they actually released an official connector, an MCP, basically to cloud code and charge APT. And it unlocked the next new planet, basically, because you can access way more functions. and LLMs are actually able to connect directly.

Publishing at Scale—and Keeping Human Oversight

And that doesn't finish even when the episodes are finished because I do create full episode plus a bunch of shorts because that's the media that actually is getting traction.

And publishing across platforms like Spotify, Apple Podcasts, YouTube, Instagram, TikTok, like you name it, I try to cover all of them. That's a nightmare, it takes forever.

But if you combine basically LLMs with Metricool and schedule and try to hit the right times it becomes way easier and then I just have the pipeline for next couple of weeks already scheduled and ready to go and I don't need to spend like you know exact time to get it the right

timing and so on and even transcription it's very powerful but it's not trained on specifics so you probably know some of like companies like antibiotics not not antibiotics, which is basically what's being read most of the time, like Deer Campus, Higgs Boson, something that are not really the popular words.

And we don't necessarily want to discuss the psychotic eye movements when it comes to robots interacting with people.

So adding the context, first of all, about the speaker to LLMs and still double checking is really necessary. So I cannot really automate everything because I tried that. Trust me, it still doesn't work. So you have to figure out where you step in where you're correct when you allow AI to move forward

The People Behind Physical AI

All right, so back to the show Basically as I mentioned My main question of the show is to ask speakers, how did you end up here basically, how did you end up doing what you do and What do you do? It's loads of stuff along LinkedIn already, but why they do it is missing and

And I already had 10 guests and the show started just in June. And, uh, you might know some of these people if you're in like tech ecosystem of Zurich, but just very briefly.

So Daniel Dobos, he spent many years in CERN. Now he is a director of innovation at Swisscom, if I'm not mistaken. And then he is organizing like AI for good and many conferences and, uh, let's say the road to Geneva for, uh, AI conference next year, he is one of the the key guys also focusing on this.

Then Katerina Portman, one of the key female voices in robotics in Europe and in Switzerland and part of an robotics woman in robotics Switzerland and so on.

Mireille, we already introduced her briefly, so she was one of the key people developing the table tennis robot and absolute tech genius, and now she just joined SoftBank. I just also found out literally yesterday. yesterday.

Nikos Kaliorakis, he's one of the co -founders of J -Floor. And if you haven't heard of J -Floor and you have a crazy tech idea, you should talk to this guy because basically they established a space, like a maker space, where you can come with your idea even before incorporating a company and get the desk free of charge. No equity, nothing. They ask literally nothing of you. They just want to get people with crazy ideas in one one space to create Kickstart and create the companies.

Then Rafael Kellner is really supporting loads of startups at Venture Lab, from incorporation to initial funding to scaling and even international expansion.

Robert McKenzie, so he helped to scale Endobotics. And now he's part -time with Ellipsis, a venture capital firm, and also had the product of Twint. And he calls himself like a Swiss army knife of tech, because he's really combining like all types of knowledge Stefan Zimmer he

helped to establish their campus that's an office space in Schlieren which have tall floors concrete floors so basically you can place big -ass robots there and that's the equipment because you cannot do that in most of the offices like in the vicinity here Katka and Anna brains behind Swiss AI weeks that got me covered so I don't need to talk much more but they have been had like very

busy time and professor yulia sanamirskaya with the zurich university of applied sciences so she is one of the key people when it comes to care robotics and neuromorphic computing so if that's the topic you're interested you must listen to her and it was a surprise that three of them were listed in top hundred women uh shortlist and actually two more in the pipeline as well but but they're not published yet, but coming soon. And I mean, that kind of showcases how strong their developments and backgrounds are.

Why Builders Do What They Do

And what I try to begin is why they do what they do. And that's where some personal stories come in, because I really get surprised, like for example,

Yuliya, she jumped from a corporation to startups, then now at university and founding startup again, and always kind of in the same direction of care robots. and I worked in that field myself. It's extremely tough area to get in and you're constantly fighting against windmills but she goes for it and when we had a chat she actually said that she had a grandma who was all that needed help and she didn't want to get you know outside people in but she didn't mind the technology to be there so it's really personal drive to do that.

For Katerina her sister has a disability and she always tries to apply AI and robotics for good. She worked at Hakoma, for example, a rehabilitation robotics company, and Women for Robotics is also targeting many of these.

Daniel Dobos, he was at CERN for many years as a particle physicist. Now he's involved in applying tech and AI for good. And one of the stories he said is how they had like a rooftop barbecue at CERN, and they discussed whether some of the packaging material they use at CERN can be used for body bags for dead bodies to preserve the body so they can actually easier and quicker identify after like national tragedies like earthquakes and so on and it's mind -blowing that they managed to pull it off and since then they actually established kind of a hackathon group to constantly try to find this fitting technology and Katka is always trying to

maximize the exposure of other creators and she grew up behind the Iron Curtain and had to fight her way out first to US, then to Europe and kind of to establish. And now what she's doing is trying to open these doors for others to make it faster for the transition.

Key Trends in Physical AI

And a couple of trends that I noticed in the first talks with the first 10 guests of TechDrive Zurich is that AI is getting a body. So we're moving towards physical AI.

AI is getting smaller and leaner when it comes to deployment.

and the trust for AI system is actually the hard part when it comes to deployment. It's not necessarily the technology itself.

The Reality Gap Between Simulation and Deployment

And the thing is that once we take AI models, put them into robotic body, into the machine, into anything, if it was trained on simulation, suddenly it has to deal with real -world physics. The carpet here is much more slippery than the tiles there, and if you fall, it's also different damage. and all of that is not quite represented in physics.

For Mirel's project, the speed was insane, like table tennis at professional level. I mean, I can't even hit like a single ball back against a lead player.

And it's really unforgiving when you deploy and when the system fails, you have to adapt quickly because otherwise the project fails.

And for Yulia, it's kind of the same when it comes to care robots. Functionality can be fine, but would you actually trust to leave your grandparents with a robot overnight that has like an arm and driving around and so on.

I mean, at the current state, I don't think anyone who dealt with the systems would actually do that. But when we see YouTube videos, it looks like very different in reality.

Trust, Adoption, and Operational Readiness

And at the same time, if you check the report that was released in spring by World Economic Forum, that's kind of the information that, you know, decision makers and executives take as an input where to invest.

best.

Two -thirds said that physical AI is game -changing, but in reality, nobody still quite knows how to develop and establish it.

Even when it comes to deployment, we see loads of videos of choreographed humanoids and so on, but when you actually bring the system into the robot, it looks like, okay, I developed the robot in the lab, I deploy it in the hospital or an industrial facility and voila, I adapted, it works.

But actually there's a lot of work behind how the facility needs to be adapted, how staff needs to be trained, how UX looks. And that is very rarely under the radar, you know, how much effort is needed to do that.

And one or three robots in the facility is great for data collection, for testing, but it's not real deployment.

like 10 to 100 robots that's exactly when we know that this probably works already and for example women in robotics run by Katerina also are establishing this robot ready framework which is exactly targeting all these

Smaller, More Efficient AI Models

steps here and in terms of the bigger models we saw LLMs VLMs VLAs and you know big multimodal models being developed by big tech companies which which get amazing results.

But when we look back at biology, we have data centers with massive power. We read about gigawatts. Our brain is like 20 to 50 watts. And can we match it with the same power? Like, no chance.

Or a fruit fly, the tiny, tiny fly, it can fly around this room and actually figure out the path and still not die on the way within first 10 seconds by using 100k neurons. The big models are 70 billion neurons. and that makes a massive difference and we still cannot scale it down same and

what we see like on the slide as you read you probably focus on the photos on like the big nice bright test text at the time you see only around 5 % of your vision is actually in focus what you really see at this moment the rest is your memory and imagination combined and AI systems are still analyzing the full scene so we need to kind of understand can we abstract to optimize this so it's really focusing on more specialized and smaller models and when it comes to trust that's also a

Security and Human-Centered Interaction

bit of the difficult part because as soon as we start trusting ai systems especially with physical body that will be a target for hackers and how do you kind of play this chicken egg problem that you know now if you your bank is hacked i mean it sucks you might lose the money but you will probably not die but if a robot fleet going crazy in the vicinity i mean that has to be considered even in the early design of creating the systems. Same in use of AI generally.

I mean, with Anna, we had the chat is here. Now, if you need to send cover letter or any document, you take five bullet points, you ask it to generate a report, you get amazing 10 -page report, which you send over. And what people do, feed it to LLM, give me five bullet points. So, I mean, do we need to play this game or do we just send five bullet points and like be done with it?

And that's something, I mean, tools are amazing, but I think we're overusing sometimes just because we can, it's free and it's easy and it makes us look very smart.

And similar with Julia on physical deployment of the robots that interact with you. Because now, for example, you look at me and so on, and if I do something like walk straight to you fast, you feel a bit intimidated. And when the big mechanical metal body moves fast towards you. You get scared and you don't want to do anything with it.

So we cover so many things in our body language and robots actually have to do the same. And they have to adapt to you personally ideally because even in different cultures, let's say the personal distance and so on, that changes a lot. And I mean the actual adoption highly depends on this.

A Path to Responsible Deployment

And there were four possible features defined in the same WEF report of physical AI.

So where we see already and our guests also spoke about that proven deployment when you have niche application in each area we already see that it's kind of getting up to speed

but we see also and we saw some tech disillusionment when there is a big hype everyone wants to get in money follows there can be hundreds of millions in investments there and suddenly like crap it doesn't work and what we see is like all the money kind of stops and everyone is back and these technologies get get halted because of maybe one or two players who did big claims, failed, and all the rest of the companies, kind of unseen companies developing there, lose their funding and can't really push forward because the trust is gone. And that is also a big issue when you kind of target this.

And divided deployment, we see that nowadays few key players from big tech companies actually developing state -of -the -art models, and we depend on them. We cannot really, you know, go around without using them easily. So there's a lot of talk about sovereign AI and it's hard to do. It's hard to compete.

But I think it's needed because we want to end up in this square when we develop AI and physical AI, where it's progressive integrated process, where basically we deploy, we test, we expand the operations and it's kind of distributed with human oversight. How we get there, we'll see. And hopefully we get into this area.

Lessons for Builders

and for builders I think you need to deploy real places teach much more than the demos so start small expand and scale mixed teams are crucial now because when you deploy the robot you can have the best researchers the best engineers but if you deploy it in cleaning service I can go there for like two days I still will not understand all the specifics and needs of like all the integrity details that make a big difference in the process so we actually need people from that domain to be part of the team or consultants or actually give us the know -how how to develop this and also have kind of a clear rules how you build trust and how you build ai in parallel to bring

that and before kind of as a takeaway message before you decide what to build always ask first first of all, why you want to build it and why you are the right person to build it. Because maybe somebody else is even better at that and that's okay. I mean, you can join them.

Closing Thoughts

So thanks a lot. If you have any questions, I thought I was the last speaker, so I thought it's pizza time, but that comes a bit later.

So a few QR codes, if you're able to scan, one is for the show on YouTube, other is to join the WhatsApp group where I try to give some insights ahead of time. And of course, LinkedIn is a must now. Thanks a lot