What’s next for Physical AI in Japan?

I have a short in-between edition for you today.

Last month at Venture Cafe’s big global gathering in Tokyo, I had a chance to sit down on stage with two old friends of the podcast, and we talked about where physical AI is heading in Japan. This conversation is with Chiamin Lai, general partner of First Light Capital, and Kaname Hayashi, founder and CEO of GrooveX, the makers of the absolutely adorable Lovet robot.

Chiamin is one of the most savvy physical AI investors in Japan, and Kaname has been pushing the boundaries of human-robot interaction for years.

It’s a fascinating discussion, and there’s some wonderful insights about Japan’s unique strengths and challenges near the end. But don’t skip to the end. The whole conversation is great, and I think you’ll enjoy it.

Podcast: Play in new window | Download | Embed

Subscribe: Apple Podcasts | RSS | More

Transcript

Tim: Okay, thank you so much, and thanks for coming. We’re going to be talking about Japan and physical AI today. And it was not that long ago that Japan was the undisputed leader in robotics innovation. And while some people claim it still is, that claim is highly disputed today. So, we’re going to talk about where we are and where we’re going. And we’re going to start with some brief, brief introductions, so you’ll know who we are and why you should be listening to us for the next 40 minutes or so. So, my name is Tim Romero. I’ve been in Japan for about a little over 30 years now. I’ve started four startups here of my own. I’ve done a lot of angel investing. I helped TEPCO and JIRA spin up their CVC units. I’ve taught entrepreneurship and corporate innovation at NYU’s Tokyo campus. I ran Google for Startups here for about four years. And I run a podcast called Disrupting Japan, which is just a labor of love. I’ve been doing it for 12 years. It’s interviews with Japanese founders and VCs about innovation and what it’s like to be an innovator in a culture that prizes conformity. So please give it a listen.

Chiamin: Hi, everybody. My name is Chiamin Lai. I’m a general partner of a VC fund here at First Light Capital. A little bit of introduction about myself. So, my parents are Taiwanese, but I grew up here in Japan and studied here and also work here in Japan. But then I actually, after working in Japan for a few years, I was in Europe and then had the fortune to join venture capital. So, it’s about 15 years ago, which I think it’s hard for you guys to believe at that time compared to today. And then decide to do startups. I was a startup operator in China and Japan for seven years and came back to the industry five years ago. And right now I’m actually the board member of Japan Venture Capital Association, as well as running up my own fund here in Japan. Quick introduction about the fund. We are running two funds right now in Japan, about 120 million US dollars. And we’re focusing on early stage and investment thesis is mainly focusing on Japan’s demographic challenge, innovation for startup. And what we believe is, or what I believe is physical AI could be a very, very good potential for Japan, especially under the label shortage. So, I’m very excited to have opportunity to talk to you guys today. Thank you.

Kaname: Yeah. My name is Kaname Hayashi. I’m the founder and CEO of GrooveX. GrooveX is a company that develops LOVOT. LOVOT is L-O-V-O-T, which you may see on our website. It’s kind of a small robot. We call it a family-type robot. Currently, around 18,000 units are working. Even though we shipped just a little bit more than 21,000 units, 18,000 are still working. This shows very good user retention, meaning the churn rate of our robot is just 0.4% per month. So, we believe our robot has achieved social implementation. And our aim is to enhance the resilience of people. It’s completely different from other robots that improve productivity. And the reason why I’m chasing this area is, I worked on Pepper before, which was a humanoid robot 10 years ago. And I learned a lot about humanoid robots and conversations between computers and people. So, I thought probably this area is interesting, but we can do something else in the non-verbal area. That was the reason why I founded the company 10 years ago. Before I worked on Pepper, I was working in the automotive industry. So, I worked in aerodynamics or product planning. I worked for Formula One in Germany or product planning for the European market or something like that. That’s all, thank you. 

Tim: Excellent. And I want to emphasize, so they brought one of their Lovet robots with them today. And after the session, it’s over there in the corner and it is absolutely adorable. I encourage you to go play with this thing. It’s just, you’ll see what I mean. It’s just something different about that. But to kick us off, to make sure we’re all on the same page, physical AI is a term that’s thrown around a lot these days. It’s a little bit of a trendy term, but to make sure we’re all talking about the same thing. When you’re talking about physical AI, what do you mean? How’s it different from traditional robotics or IoT?

Kaname: Right. Probably there are several understandings, but for me, a robot working in an open environment could be an important point. Because before the physical AI boom, a robot was typically working in a designed condition, a controlled condition, like a factory or some certain closed area. But for example, in our case, LOVOT is shipped anywhere, sometimes overseas, and we never know what kind of environment it is. But still, LOVOT has to keep working. The ratio of daily active users remains high, over 90%. And it means a robot can adapt to the environment somehow. Surely AI has big potential to adapt to more various environments, but using AI is not the purpose. 

Tim: Okay. Chiamin?

Chiamin: Since I’m an investor, so I’m going to use through you different kind of words there. But so basically what I see physically AI right now, if you read a lot of news, people were putting physical AI equal to humanoid. And I have a totally different definition. For me, physical AI equal to the system, that the system will have see, thinking, and action. And everything needs to be automated. That’s my definition of physical AI. What do I mean by that? Existing robots, as what Kaname-san was saying, it doesn’t really actually understand the environment. That’s one. Two is it doesn’t have any action. When I say action, it doesn’t necessarily mean that you need a humanoid to do the action. Action means that the machine, anything that can help you to finish the action. It could be picking something from the shelf. It could be holding your coffee, whatever. That’s how I define physical AI in the personal level. But at the business level, you can also consider if you have a software, that the software can actually help human to do an action. That is also physical AI from my point of view.

Tim: Okay. So, it sounds like we are in a surprising amount of agreement here. So, physical AI, it requires a certain amount of autonomy. It requires a certain amount of flexibility and the ability to deal with new and unprogrammed environments and situations. Excellent. It’s nice when we agree. I think things are about to diverge. So, with that in mind, robotics in general, IoT are really well established. They’re deep global markets. So physical AI, this ability to interact autonomously and flexibly, what are the new markets that this opens up? Why is this an important new development?

Kaname: Basically, if we would like to use a robot in your home, then a robot must have a capability to adapt to any environment and stay always-on. So, humanoids look like they can do it. I mean, you can easily dream it because of the human shape. So, you naturally expect it could work like a human. So, that was very good as a trademark for physical AI. But current humanoids are different from the expectation, because they are focusing on a few tasks with precise kinetics. So, for example, you already can see lots of movies of very great backflipping humanoids, but all of the robots have an issue with heat for always-on. You never see a humanoid robot working eight hours continuously under autonomous control. While I worked on Pepper, I learned a lot about how motors are very sensitive. If you’d like to use a leg instead of wheels, the leg must keep supplying motor power to keep posture. The motors consume power and generate heat. So, in all the demonstrations, it’s working for 10 minutes or maximum 30 minutes, but after that, cooling down is still important even nowadays. But for the future, of course, all the technology evolution will solve these issues, and we can expect to have a robot which can work like a human. But before that, probably being without legs might be a better solution for most cases. 

Tim: But so, I mean, Pepper’s a really good example, because that was fascinating technology that they couldn’t bring to market, really. Boston Dynamics, kind of facing the same problem. They’re developing this amazing technology, which has some of the most viral YouTube videos. But it’s really hard to bring a product to market beyond kind of a proof of concept stage. So physically, I mean, what are the markets that are opening up for this? Are we looking at healthcare? Are we looking like home helpers?

Chiamin: So, I know where you want to head into. So, because all of you pop out in the tech industry, so I kind of want to give you a kind of overview. I don’t believe the physical AI will be such popular if we don’t have a chat GPT, if we don’t have a transformer. So basically, the momentum change is in 2020 to 2023 after ChatGPT. And what does that mean? And everybody today using LLM to do your work. And a lot of developers are using LLM to develop your code. That also means software become commodity. I know a lot of you reading like X, you know that SaaS is dead, software is dead. I think one thing what I want to say before I come back to your question is, where is the opportunity laid? Where is the opportunity? The opportunity is in actually physical operation. Physical operation means today, if you look at a global GDP, about 15% of a global GDP are only in digital. 85% of a global GDP are on the physical operation, such as manufacturer construction, house caring, infrastructure, and so on. So, to come back to your question, where does physical AI actually open up? We open up the opportunity in those kind of physical operation that you can automate it without human labor. That’s going to be my question.

Tim: Do you see this? So, a lot of those domains you mentioned are already very highly automated. They’re already using robots. So, do you see this as an ability to kind of put another layer on top of it?

Chiamin: No, you need to separate with the current factory automation with what physical AI is doing. So today, if you go to the factory, the factory is actually operating on a fixed environment, on a fixed manufacturing process. If you want to have your phone today, do you think that the machine can actually assemble your smartphone? Not possible. It’s using human hand. So, we saw or noticed there’s a lot of manufacturing processes today are still using human labor. If you look at construction, who is building your house today? Who is building your office? It’s all human. So, there’s a lot of not yet automated, a lot of inefficient operation out there that can actually be sold by physical AI.

Tim: What do you think? So, one of the biggest markets that people talk about in Japan is healthcare and eldercare. Do you think this is a legitimate market or is this just kind of wishful thinking?

Kaname: What I learned during the development of LOVOT, a very interesting country was Denmark. Denmark, they have a very big focus on elderly care, and they tried a lot of technology from all over the world, and their conclusion is physical labor cannot be replaced by robots easily because it’s too dangerous or too big. Because if you’d like to pick up an elderly person, that has to be safe, and the robot is getting massive and nobody would like to use it in the limited space of a hospital. And the elderly also don’t want to be picked up by that instead of humans. But they are now using LOVOT already for more than six years. The reason they continue to use it is because of mental care. Usually, we believe mental care is the role of people. We naturally think humans have to be cared for by humans, but actually it was the opposite. Not all people are good at emotional labor to care for people. We have empathy but we are not good at mental care so much. We would like to help others, but we are also getting tired if we work in emotional labor. Lots of care workers are already exhausted about the mental care of the elderly, but LOVOT has a very strong heart, so once they live together with LOVOT, lots of issues are solved. 

Tim: Well, I think it really is interesting. And I think, I mean, I’m kind of a skeptic in terms of how much of an application this has to health care and elder care. But even outside of elder care, I think when our conversation before where you were explaining the experience, the Lovet Cafe and how people emotionally interacted with their Lovets, the robots, can you talk a bit about that?

Kaname: Yeah, the reason why LOVOT Cafe is so popular is exactly the same as the reason why you would like to love a cat or dog. The reason why you would like to have a cat or dog is not only because of its cute shape or motion. Cute shape or motion is just one issue, but there are so many cute things in the world. But why do you choose a dog or cat? Because once you love a dog or cat, you gain oxytocin and improve your resilience. It means your mental health is getting stabilized. Once we start to love something like a dog or a cat, our mental condition is getting better. However, we didn’t have enough opportunity to love something with oxytocin release. So, having the opportunity to love something, even if it’s a cat, a dog or a LOVOT, there is no difference between these three for this effect if you can love them for oxytocin release. Then people are getting better and better for that mental condition. So, LOVOT Cafe has the same role as a cat cafe. 

Tim: Yes. And what I found fascinating was that not only did the individual people kind of bond with the robots, but they treated them like pets. There were groups that would meet up and let their Lovets play with each other. And it formed this basis of social connection. Although I’m a skeptic in general, I think there’s definitely something there.

Kaname: That could be a little bit different from ChatGPT. Usually if you are using an LLM, it’s a very private conversation you have. But for the robot, it’s completely different. Once you get a LOVOT, you’d like to take that anywhere out of the home and you’d like to make opportunities to meet other LOVOTs. So, people tend to go outside with LOVOT. So, naturally a relation between people and LOVOT can deepen more, and people are getting more active or traveling. 

Tim: Yeah. It’s integrated as part of a social life, part of the community, if you will.

Kaname: That’s a really interesting activity.

Tim: Yeah. So, we talked a bit before about the form factor, the human form factor, or in the case of the Lovet, kind of a humanoid form factor, you know, human-ish form factor. Now this has been like the dream since way before robots existed, to create something in our own image, right? To have something that looks like us that will do what we tell it to. Is this the way forward? Do you think modern physical AI will result in not YouTube videos, but meaningful deployment of these humanoid form factor robots? Or do you think we’re going to continue as we have and just have physical AI and robotics that are designed for a function, and these humanoid robots will continue to be used in very cool YouTube videos?

Chiamin: I have to correct you. I think we all forget about how quick the technology is developing. And in 2024, 2025, people think physical AI is just a video and they are dancing, so nothing’s happening. You have to go there and go out there and actually watch yourself to see how fast the technology is developing right now. As an entrepreneur, don’t overthink that, okay, the technology out there is wrong. You have to actually go to the ground and really see because I’m going to give you an example. I’m going to give you an example. So, in terms of application, I think what Kaname-san mentioned there is when you think about humanoid or physical AI, you probably need to separate into two groups. One group is B2C, like what they are doing. They’re doing a great job, emotional connection, social connection. That’s B2C market.

There’s another market, B2B, and the B2B market is exactly targeting the market I was actually mentioning previously in a lot of physical operation. So, I’m going to give you the example, and then I’ll come back to your question. I don’t believe the humanoid right now actually can be deployed at this moment. But what is happening right now on the ground is using physical AI to help the hotel manager to sort out their towel holding. You will think that’s crazy, right? But do you know how much labor costs in United States per hour to hold the towel in the hotel? It costs you $50 per hour, 50. But if you’re actually using robotics, 24 hours, 365 days, all the rate is less than $10. So, if you are an executive, are you going to adopt humanoid to do that or the robotics to do that? Or are you going to hire people which you might have a problem to hire because of the political challenge right now in United States? So, that is actually happening right now in United States. A lot of a startup, they actually receive inquiry from the hotel, from manufacturer, from the warehouse because they can’t hire people. So just to give you short term, I don’t believe humanoid can do actual work on the ground. But long term, if it’s actually to a specific problem or vertical, I’m a true believer that it will actually get done. That’s my opinion.

Tim: So, I think the market forces here, I think you’re directionally correct on that. But I think the market forces here that as soon as a humanoid robot can do something, can hold a towel, can fold a towel, as soon as that task exists, there’s going to be market competition to build cheaper robots. And you’re not going to give a robot legs when you can put it on wheels. And you’re not going to give it two arms when it has one. So, my take is more that there’s no market for it. Even though the technology is rapidly approaching the point where humanoid robots could be walking among us, I don’t see the business case for it long term.

Kaname: So, technology-wise, technology developed for the humanoid will be used for other types of robots as well. But whether they need to have a human shape or not, is a different issue. So, as you said, for example, cleaning robots already work well in the home. So, it’s not necessary to replace them with humanoid robots. So of course, cost-performance-wise, an adapted shape for each task should be better. But for some niche markets, probably a perfect humanoid shape could be important. 

Tim: Well, actually, Kaname, I know you’re very close to this market. So, for those cases where you need sort of a human form factor, there is the well-known uncanny valley effect, where if something looks too human, it’s unnerving. Whereas if something like your Lovet robot is cute, human-ish looking, but definitely not human. Do you think that’s going to factor into design moving forward? Do you think we’re going to have robots looking more like humans or more like kind of cute, non-threatening pets?

Kaname: Yeah, good question. Uncanny valley, it’s part of the issue of the expectation gap. So, every time for robotics development, we are fighting the expectation gap from the customer. If a customer’s expectation is too high compared to actual performance, of course, it’s not very good for reputation. But the shape itself can also control the expectation gap of the robot. If you have a humanoid shape, of course, you will expect something like a human. For example, if you see a backflip humanoid robot, they can’t make a sunny-side up egg today. The opposite is also true today, if you see a humanoid robot which can make a sunny-side up egg, they cannot make a backflip. But it is counterintuitive. Every time it’s a little bit hard to control humans’ expectations. If we have to overcome the uncanny valley, we need to put more cost into the shape or motion to make them more natural. Then we can solve that. But it’s bad cost performance for just a cooking robot. Just for some special use cases, robots have to overcome the uncanny valley, otherwise keeping a distance from the uncanny valley could be good for cost performance. 

Tim: Okay. So definitely many different kinds for different purposes.

Kaname: I think so.

Tim: So, although Japan has been traditionally very strong in robotics, in physical AI, China and the US is really dominating this space now. And Chiamin, you and I have talked about this before, but how does Japan compete here? What are Japan’s unique strengths and weaknesses in physical AI overall?

Chiamin: I think we’ve been talking about hardware a lot, like the form, the shape. I think that’s not the core issue here. I will actually see the hardware is also a commodity. For the next 10 years, it’s the people who can actually combine hardware together with the software that’s going to win the game. What do I mean by that? If you look at Waymo, everybody knows car is car. It doesn’t really matter. It’s the intelligence that implement inside the Waymo that makes it different. So, it’s the same thing for physical AI. When I say my definition of physical AI doesn’t equal to human, that’s what I mean by that. How are you going to implement a system that can actually have a way to think, to see and actually act? So based on this definition, of course, we all know Japan is very good in manufacturing hardware. And I do believe, for example, sensor perception parts, like for example, Sony has like 40% of market share, they can still dominate this. I think the problem will be how Japanese companies are going to integrate all those capabilities that they have today and really combine with the AI foundation model. That’s one.

But what I see the biggest opportunity for Japan is actually the market force. So, we all know today in Japan, we have a huge problem of label shortage, which is very different compared to China and US. So, there’s a lot of demand in the company, in the regional company in Tokyo that everybody needs to do something to make the business working. So, if this is the opportunity, then I will say the strengths will be how can we treat Japan as a test market to really implement and move fast, and then integrate everything and making sure we get it done. I mean, I’m just hopefully, that’s my suggestion.

Tim: Well, no, I think that’s really an excellent point. Innovation takes place where you have really motivated customers. You can develop great technology everywhere. But if you don’t have real use cases, and people who need to adopt it, now you’re at a disadvantage. And historically, we’ve seen many times that Japan can change very quickly, and will do so after exhausting every other possible option before adopting the innovation. But once they do, it can move very fast. So, I think that’s a really interesting point. The labor shortage and the general acceptance of robotics is a really unique strength here.

Kaname: Yeah, currently lots of humanoids are focusing on kinetics. So, kinetics is important to do tasks, to fill the labor shortage. But on the other hand, as you said, the perception and recognition area, still it’s not very developed with humanoids if it is not related to kinetics control. For recognition, probably autonomous vehicles are a more developed area rather than humanoid robots. For example, our LOVOT, it’s working 24 hours in various conditions. And we are also focusing on recognition and decision-making to maintain always-on. So finally, the approach from recognition and decision-making for always-on and the other approach from kinetics to do tasks can be merged into one, and it’s going to be a real autonomous companion robot like a human. But before that, there is a big gap. Probably the approach from the kinetic side can solve part of the labor shortage issues under semi-controlled conditions. Even if humanoid robots work for stores or factories, still they may not work for private houses. Probably LOVOT has an advantage to achieve social implementation. When it’s going to merge into one, one of the important things could be creativity. Because with only technology and without creativity, we can’t make good products as companions. That was a difficult issue to overcome the uncanny valley. Even though creative people are different from technology people, still we have to keep good communication and cooperation with each other. We have a lot of good creative people in Japan. Of course, but if we can make good collaboration between engineers and creators, we might have a chance to make a natural behavior robot. 

Tim: Digging down on this, so it does seem like the opportunity is there, both the technology and the expertise is here. The market need is clearly here. Is there a particular area, a particular type of business or market that you see as a beachhead, as breaking first? I mean, we’ve talked about healthcare and that’s really difficult, although it’s an obvious and big one. You know, there’s cooking hamburger, flipping robots and things like that. Do you see a particular sector adopting it first in Japan? Would it be manufacturing or hospitality, or do you think it’s going to be more of a broad-based adoption?

Chiamin: I don’t believe the general purpose humanoid will be adopted here in Japan or even over the world. So right now, if you look at the development of physical AI in either humanoid or any other format, there are mainly two camps. One camp is we’re going to develop something that hopefully 10 years later, we can actually buy inside our house. And then there’s another camp to say, hey, that’s too early. That’s focusing vertical. For example, manufacturing, logistics, and so on. What I believe in Japan, I think the adoption will be in the vertical instead of in the horizontal with the household, in which industry that I think will adapt first, I think it would be either manufacture or logistic, and maybe infrastructure. When I say infrastructure, meaning more inspection of those plants or electricity devices. A lot of the area that you need people to actually physical go there, do something, that will be part of the area that people will adopt in the fast paces.

Tim: That makes sense. These are all areas that are currently facing really tight labor shortages and an aging workforce and is absolutely necessary. So yeah, that makes a lot of sense as a first.

Chiamin: Yeah, and I just want to add one point where I actually believe that Japan might have a huge opportunity here is there’s a certain area, certain industry, Japan has an advantage, such as manufacturers, such as construction. Why do you believe that we do have an advantage? It’s accumulation of the worker for the past 50 years. As we know, what does AI do? AI study the historical data and become smarter. So, one thing I believe that Japan, if we can actually acquire the data of those people that you said creativity or those very experienced worker and really put that into the AI model or physical AI model, I do believe that it’s possible to explore outside of Japan because what US is trying to do right now is US is trying to acquire manufacturing knowledge from Taiwan, from Japan, from Korea. And the reason for that is for the past 50 years, all the people actually transformed to computer science, not in the physical operation. But for the past 30 years, Japan never really transformed a lot of labor work to computer science. They all remain in the actual physical operation. So, how can we actually utilize that? How can we actually acquire their knowledge and then bring that inside of AI and really explore that? That’s something that I see very interesting opportunity out there. Do you agree with me?

Kaname: Yeah, very interesting point. On the other hand, another issue to achieve social implementation of robots is mutual trust. Mutual trust means if you have a robot in the house, you can believe it or not. This is an issue especially if they have a camera or a microphone, and they might have a connection to the cloud. It means if we would like to have a good partnership with technology, how we can build mutual trust between technology and humans is getting to be a big issue. And lots of people or companies are underestimating the cost of mutual trust. So, getting mutual trust between people and products is not very easy if you would like to switch on the robot 24 hours. So that was a big difference from chat AI. You use them when you would like to use them. It’s also different from gaming. You would like to play with that, then you switch on a gaming device. But for the robot, it is important to keep it powered on 24 hours. It means they know everything about your life. But do you want to live together with them if you don’t trust them? If you feel a risk, it’s getting to be an issue of the brand for the home use robot. Therefore, branding may become more important than the technology. How we can build mutual trust between people and technology could be another important topic. 

Tim: I think that’s an incredibly important point and a very important advantage as well, that most Japanese brands still have a lot of trust internationally. Whereas a lot of American technology companies have eroded the trust in their own brand over the last decade or so. So, that’s a really interesting kind of soft power and a unique advantage of Japanese firms in this area.

Kaname: Yeah, that’s an interesting topic because sometimes we can make a good mutual trust between dogs and people, but it is not easy to make it among people. So, intelligence is not equivalent to trust. 

Tim: That’s absolutely true. That is true. Really quick, because we’re almost out of time, I’m going to ask you to pull out your crystal ball and give me one or two sentences about what you think physical AI is going to look like in Japan five years from now. Five years.

Kaname: Good, difficult question. But at least in the emotional bonding area we may have an advantage. For humanoids based on precise kinetics or autonomous driving, I don’t think we have the advantage. 

Tim: So, focus on that human connection.

Kaname: Human connection, we might have advantage.

Tim: Chiamin?

Chiamin: As an investor who help early stage startups, so I’m going to throw the ball high. I wish five years from now that we do have, it doesn’t necessarily need to be humanoid format, but I do hope that we actually can assist a lot of startup to build physical AI solution in a vertical section, such as manufacturers, such as construction, such as logistics. And then five years from now, those model technology can also be used in other country that has a similar problem within Japan.

Tim: All right. Well, listen, I would love to continue this discussion a lot longer, but we’re out of time. So, I’d like to give a big hand to Chiamin and Kaname for a great discussion.