Loading summary
Transcript86 lines
- [00:00]
A
So the robot was transmitting data to servers in China at regular intervals during normal operation. Audio from the microphones, video from the cameras, GPS coordinates, all the torque readings, spatial mappings from the LiDAR, everything.
- [00:16]
B
One of the big questions in thinking about the future impact of AI is robotics. When people talk about a future where AI has a big impact on the labor market, for example, it really matters whether robots will have the ability to do physical jobs that human workers do. And that depends on whether we'll have robots with human life capabilities. And it also matters where these innovations will happen. If they happen in the United States, that suggests a different future for the world than if it happens in another country such as China. And so I've got the perfect person to talk through these issues with me. Divyan Kaushik is vice president at Beacon Global Strategies and he has a new substack called the Body Problem that I've really been enjoying, where he goes into a lot of details, done a lot of in depth research about robotics, about the difference between the US and China, but some of the regulatory questions and so, Divyaj, thank you for coming on AI Summer.
- [01:04]
A
Thanks for having me. I think it's going to be a great conversation. I hope so.
- [01:09]
B
I said, you're vice president of Beacon Global Strategies. Can you explain for people what is Beacon Global Strategies and what do you do as a vice president?
- [01:17]
A
Yeah, I think the best way to explain it is Beacon as a national security advisory firm. The core thesis is that national security and foreign policy are shared objectives between the government and private sector. And how do we get the private sector to do a better job at furthering US national security and foreign policy objectives? So we help C suites think through geopolitical national security questions. You could think of it as like the private sector's national security advisor. So I've been thinking a lot about what issues American private sector will have to worry about, or will have to at least start thinking about answers to AI policy, say five to 10 years from now that we will questions that we will encounter then and led me to this void where nobody was really doing actually good in depth work on what it means for robotics policy where AI does indeed get a body. So that led me to start the substack because I thought it would be helpful to reason in public and think about these issues, because I don't believe that even I have all the answers. I'm just trying to think through what are the questions that we need answers to. And that brings us here.
- [02:47]
B
So concretely, you've got a company that maybe makes robots or maybe expects robots to be effective business somehow in the future. And they're thinking about what policies would we like to see and what can we do as a company to prepare for possible policy changes? And they hire you to give them advice about both what they can do and what they can ask policymakers to do.
- [03:05]
A
That's one thing. But there are other aspects which are like supply chains, export controls, how do we think through. How do we think through compliance? How do we work with international partners? How do we think about diversification of our supply chains? What impacts could potential tariffs or export controls have on us or.
- [03:25]
B
Or in Iran, for example.
- [03:27]
A
Or in Iran, like I. The amount of time I've spent talking about Iran over the last month or two. Yeah, I. It's like how some people have gotten fed up of the word magnets. I'm like, I do not want to talk about Iran again.
- [03:46]
B
Well, I do not think I have any questions for Iran, so that'll be good. Talk to me about how you get into that. I mean, this sounds like a somewhat esoteric set of topics to be an expert in. And you have a computer science background, right?
- [03:59]
A
Yeah. So I got my PhD in machine learning from Carnegie Mellon in 2022, where my dissertation was on auto distribution robustness of NLP models, which today we know them as LLMs, large language models. And during that time I was helping out with CHIPS and Science act and got to do a lot of policy work during my CME days with members on the Hill and in the administrations. Both Trump won and Biden administration.
- [04:34]
B
But unpack that a little bit when you were helping out with the CHIPS Act. I mean, not that many grad students end up, quote unquote, helping out with major legislation. How did that come about?
- [04:45]
A
Yeah, I think it's once you make yourself useful like you make yourself, that's a trap, you get more sucked in. I think I got into this in 2017 when there was the Tax Cuts and Jobs act that was being discussed and it would have taxed graduate student tuition waivers as income. That was an issue that was going to personally impact me. So that's how I got involved through Carnegie Mellon's graduate student government at the time. But Senator Toomey trying to fix that and we did. Through that I built a lot of relationships and I realized that people were craving technical expertise on some of these issues where they were not getting access to it on the Hill outside of just getting to talk to the government affairs people for all the companies. So I started talking to the committee staff who were thinking about these issues at the time on how science or Send Commerce and Start a to help chips and science as a result of that, including the 20,000 names that went through Endless Frontiers or US Innovation and Competition act or NSF for the future ACT or America Competes Act. I don't remember any of the other names that we ended up thinking about. But yeah, that was a big, big effort that made me just more interested in policy. And I think my PhD advisor was very kind enough to allow me to spend a lot more time doing this. And he would say that policy was always my mistress. One of the things I think is
- [06:36]
B
remarkable at the career you've had so far is you've got pretty good connection on both sides of the political aisle, which I think is pretty unusual these days. I mean, often you have somebody who's like a Democratic operative or Republican operative and it sounds like, and you mentioned that you work with, with people in the Tax Custom Jobs act and more recently with the Biden administration. And I think you still have some strong Republican connections. I know that seems valuable, so kudos for that. Yeah. So, so let's, let's talk about robotics. So there's been a lot of progress. I mean you were saying that this was, was something that you were researching. What's the current state of the models that we use to guide robots? And how has that changed since you got your PhD in 2022?
- [07:20]
A
Yeah, I think there are a lot of things out there and it's very important to walk through the main categories of robots that we see. There are the industrial robots, the old mature core, which is factory robots that people imagine. The arms building cars, whether it's painting parts, assembling electronics, whatnot. They're all operating in structured environments. They are factories, cages, work cells, whatnot, assembly lines. Right. And these are extremely important. But they're not the intelligent in the AI sense. They are precisely pre programmed machines. They're doing repeatable motions. That is the most widely deployed category that we see today. That's. And the reason why that is important is the biggest robotic story. And even though you may read all these stories about humanoids, whether it's a humanoid joining a Buddhist or becoming a Buddhist or running a half marathon or whatnot, the biggest story is still not humanoids. It's the boring, productive industrial automation. And then you've got the collaborative robots, which are the robotic arms that are designed to work near people. They're working at often at lower speed or force than traditional industrial robots. Whether it's machine tending, packaging, light assembly, warehouse, small batch manufacturing. And that just is kind of some degree of human proximity. They are a bridge between fenced off industrial automation and the robots that share space with humans. Very similar to warehouse and logistical robots. And then you start to get to all these medical, surgical and whatnot. Some people also consider drones and AVs as part of robotics, which conceptually you can see how they are just like instead of walking and running, they're flying or they have wheels that operate at high speed. But then you've got what you're seeing today with general purpose humanoid robots where the core idea is that the world is built for human bodies. It's stairs, doors, handles, whatnot. And a humanoid form factor is an attempt to use the existing human infrastructure rather than redesign the world around robots. And to do that, people are trying to use different forms, whether it's designing newer models. I think some of the companies, I think we just saw pretty good demo from a French company that came on Twitter recently, but also American companies like Physical Intelligence and Skilled AI. They're doing some really good work in designing the models. These models started off as matrix manipulators back in a couple decades ago to Vision language. Action models And VLA models are basically like they're kind of your LLM models. They follow similar scaling laws. The data is different. Yes, you can use the video data, you can use simulation, you can use a lot of things. But at the end of the day, what you need is proprioceptive data. You need people wearing exoskeletons collecting data, doing tasks, or teleoperators operating robotic arms and collecting that forced torque data that you need a robot to learn from. So each category comes with different governance problems. Each category comes with different liability problems and different kinds of issues around when you think of AI governance. So the simplest way I would say is these are different form factors and different kinds of brains that are in those form factors that are trying to sense the world and physically act on it.
- [11:23]
B
So you mentioned VLA models, which is vision language action models. Let's unpack that a little bit. That originally we had LLMs which are models that you have some text and you try to predict the next token in the text. And it turns out when you do that, you can get models that can, that can produce realistic human speech. And then we had vision language models, which was the same idea, but you could, in addition to having tokens represent words, you could have tokens that represent pictures. And so you can have LLMs that can understand images and now you got vision language action models which are vision language models that also have actions which are tokens representing robot actions. Is that right?
- [12:06]
A
That is the right way to think of this. Which is why the action data is very important. And what comes with that action data is from a video of me holding this bottle. I can see how to lift a bottle or whatnot, or how to place it down. But what I don't understand from the video is how much pressure to apply when holding this bottle versus a wine glass.
- [12:31]
B
So the actions are going to be move right arm 3 degrees or tighten gripper or things like that. And so one of the challenges with any LLM or any of these systems is that you need training data. So with text, if you have a sequence of tokens in a book or a newspaper article or something and you're trying to predict the next action, and so naively what you'd like is like the same thing. You'd like a bunch of examples of robots doing things so you can do the same thing. You can say, well, the robot saw this scene and then the action it took was this action. And so we want to train a model to do that action in that situation. Walk me through more why that's harder to do for robots versus for language models.
- [13:17]
A
Yeah, I think the reason it's hard, there are actually more than one reason why it's hard. 1. I think there are a couple of theories that I've heard, especially from a couple of Chinese founders who make a very specific case against VLA models. They say that VLA is very body specific. You change the embodiment, you have to recollect the data for that embodiment, you have to retrain because every form factor is different. Quadrupeds are very different than humanoids and you need very different. Now some of the labs are trying to do cross embodiment trainings separate from that. It's just like the core argument that the VLA critiques are around. They can only learn from robot generated data and there aren't enough robots to generate the data at scale that matters. Unlike say autonomous vehicles where cars existed at scale before anyone needed their data. VLA ignores what is in the digital knowledge base to some extent. Right. It somewhat ignores text, physics, reasoning, it ignores books. That should be part of how a robot understands the world.
- [14:35]
B
Now isn't that supposed to be in the pre training? If you take an LLM, you pre train in that.
- [14:39]
A
Exactly. And then.
- [14:40]
B
Right.
- [14:41]
A
And so some companies are trying to challenge that and trying to show how vision language, action Models with pre training could work. VLA has been hard for these reasons so far. But with pre training where you are training one first, like you're pre training the models as a vision language model and then appending action to it. I think several of these companies are trying to solve for the problem where they can train VLA models on multiple robot IDs. The challenge is usually then adapting VLAS to your robot. You can fine tune on small demo sets. There are a ton of people who are trying to use human data to transfer to VLAs. There is also just like to your point on pre training, it's not true that VLAS don't retain any data from pre training, even though I've heard that point repeatedly made by Chinese founders that they don't. But yes, they lose a lot of the generalization and people are working on different things on how high or low level vlas. People are working on different things like high or low level VLAs. And those try to retain more of the VLM knowledge, the vision language model knowledge. Almost everything that people are proposing to do with world models, for instance, and robotics, you can't do with VLAs. Incidentally. VLAs is just like a particular kind of model that starts from a VLM backbone to produce action for robots. Now it could include any number of training or architectural methods that depends every company to company. But pre trained models I do think are going to play a huge part. Word models are a very exciting direction since they do tend to learn more of this physical common sense that VLMs often lack. But I do think VLMs will still play a huge role in robotics as long as we can. You know, as long as I think we care about robots interacting with people in any way, we need a language component. VLM pre training will play a huge role.
- [16:47]
B
Okay, so VLA is one leading approach to building a modern deep learning based AI system for robots. But it's not the only approach.
- [16:56]
A
It's not the only approach I've heard from people that several Chinese companies are thinking of while they currently are employing VLAs, they're thinking of newer approaches around world models similar to how Jan Lecun is thinking about LLMs.
- [17:14]
B
So for any of these models, these approaches though, you need data.
- [17:18]
A
You need a lot of data.
- [17:19]
B
And like with the Internet, with LLMs you could train them, you've got all this pre existing text, but we don't have a big database of robots doing things that we can just train on the way. We have a big database of text. And so we had to collect them. Which is why you see companies having people put on suits or virtual reality goggles or hand sensors or whatever, and having them do the thing so that then the robot can. You can train the robot on what you're doing. And something you've written about that I think is interesting is we're starting to see companies do this at scale, and particularly in China. You have a lot of people, a lot of companies doing this. Is that right?
- [17:53]
A
Yeah. So last year, China announced this initiative where the state was creating 40 new factories for data collection where humans just perform the same task hundreds of times wearing exoskeletons. At First, a robotics executive mentioned it to me last year and I just, like, ignored it as a throwaway detail. But then I came back to this after I interacted with a couple more people who are building robots in China. And it's pretty fascinating what they're trying to do. Two dozen of these factories came operational last year. The remainder of them should be operational this year, I think. And since then, I've talked to a couple of companies trying to do data collection in the United States. It's just really hard. It's really expensive. The robotic bodies are expensive to collect that data. I was at a robotics company last week trying to teleoperate an arm just to get a feel of how hard it is. It is really hard for. Unless you are really trained in teleoperation, it's just very hard to collect that data successfully. The successful examples, because you can't just. You can't just collect the failures. You have to, like. One executive told me during that visit that failures are good if you can recover from the failure and you're collecting that recovery data.
- [19:14]
B
When you were doing that, what specifically what task were you trying to do?
- [19:16]
A
I was trying to put three fasteners, three screws in a plastic bag, and just like Ziploc it using the robotic arm. Similar to how you would think of an Amazon factory, like Amazon warehouse, like, packaging stuff that is on the shelves for fulfilling a consumer order.
- [19:38]
B
And was the main thing that was hard to do, the kind of spatial reasoning, given the cameras you had. Or is it like you didn't have force feedback when you actually picked it up? Or what was the.
- [19:46]
A
No, you had everything. Except. I think it's just really hard to get to a place where you know, because you're operating a leader and a follower arm is following you, there's just a disconnect from that until you get like, fully trained on how to get. You're not directly getting Force feedback from the leader because the force is happening in the follower arm. So it's just like really hard sometimes when you're like, okay, I have to pull this plastic bag from a pile of plastic bags and I'm trying to pinch it. I just don't know like whether I applied enough force or not.
- [20:28]
B
And is your sense once if you did it for a week or a month that you get to the point where you can pretty much do it as well as you can yourself or it still be like kind of cumbersome?
- [20:35]
A
I feel like, yeah, I could do it if I was training on it for a week. I feel like I could be better a month. I could definitely do it. But I've seen a lot of companies that try to collect this data in factories in Mexico and India a lot more just because how hard it is to collect this in the United States, how expensive it is right now. And this whole exercise for me was just mind blowing. One, it's just like a whole. It rearranges neurons in your head when you try something yourself rather than just reading about it or watching videos. The other thing was the economics. You may have 30 or 40 robotic arms in San Francisco, but you can do 10x that data collection in Mexico or India. And so a lot of American companies are trying to do that now.
- [21:34]
B
One of the things that was crucial for the success of LLMs, as I mentioned before, is that you had all this Internet data you could scrape. It was kind of this public resource that anybody who had an idea for a better model could use without having to pay the kind of full retail price for all that content. You can debate whether that was good or bad, but clearly that was helped LLMs get off the ground more quickly. And it seems plausible that robot data could have a similar kind of public goods element that if there was a big database of robot manipulation data of some kind, that that could then make it much easier for a startup or a research lab to do robotics progress. And so you've written a piece about a possible role for the US government for this. And in particular, one of the things that's interesting that the US Military interesting about the US Military is the US Military has this immensely wide range of environments and tasks that they operate in. Talk to me a little bit about the role you could imagine the US military playing in this.
- [22:30]
A
I think this is really important because in March the Navy entered into this contract with Gecko Robotics. Or for robots that will climb a ship's hull to identify hull damage. That's very perception based and that is very important. But it's also like the distance between perception and force is long way. Like we have a long way to go from here there. Because contact rich tasks are really hard that require a lot of data. As we were talking about, I'll give you an example. Like I think pi0 paper that was just like last year or something. You are using only automated training data simulations, basically no manual close contact demos. And same was an easy insert paper that went viral last year. It was if you're inserting a type C cable within 1cm of the socket, the model was 100% accurate. Within 5 millimeters, the accuracy dropped to 60%. And for the final insertion it dropped to 30%. So I like to term this the last 5 millimeter problem.
- [23:56]
B
So it could get it close, but then you have to be exactly lined up. And so it'd like kind of like slightly off center and then it wouldn't go in the way it's supposed to.
- [24:03]
A
It wouldn't go in. Exactly. And so this pattern is like consistent across every frontier model that has published both contact light, contact range benchmarks. Performance just collapses in the final phase of manipulation. So when I say that the Pentagon has a lot of diversity in data, like simulation closes the gap. We've seen that, but it just does not get us to the last mile. And pi0 really showed this to us. Where when Physical intelligence released this paper where they had a Force VLA comparison, they had a PI zero based model which was trained without force data on a task like the jump from the performance of that model on tasks to performance of the model with say force and torque data like force vla. That was another model that they had trained. The jump was closer to like 30 to 40%, sometimes even 80%. So this contact gap is like really important. So PI 0.6, for instance, it's trained on thousands of hours of diverse robot data, handles, shirts, dishes, near human reliability. The moment the task demands sustained force control against say a stiff environment, whether it's assembling a box, it's inserting a connector, the performance just drops by 4 to 5x. Because contact dynamics require control loops that are running at frequencies that current vision language, action architectures that we've been talking about, they just cannot reach. The VLA architectures do not reach them.
- [25:37]
B
So unpack that. You said sustained force. Is that picking something up, picking something
- [25:41]
A
up, pushing something against something, like keeping it held tight when you're applying the force, the forward force. And so as the contacts get stiffer, like controllers need to run faster, to stay stable. That becomes a hardware problem too. This physics forces an architectural split at
- [26:03]
B
the end of the day because holding something is kind of a complex feedback loop, right? You want to put enough force, but not too much, because if you do too much, it'll move it or whatever. And we don't think about it because our hands are, you know, our hands and our bodies are like very. We have years of practice, but like it's actually a hard problem.
- [26:18]
A
Exactly. And so in a slow loop, say a VLA model, you're operating at, say your motors are operating at say 20-50Hz frequency. You're doing grasp selection or something, right? But in fast loop, where your motors are running at 500-1000Hz, you have force impedance controllers. You're thinking of contact stability, you're thinking of reactive force adjustment. What happens when I've picked up a screw but now I'm trying to insert it and I'm getting reactive force? How do I need to adjust my force? You need to handle mode switch handling so things become harder. The Pentagon, to your earlier point, it has so much diverse data, so much diversity that exists in the environments at the Pentagon. The shipyards are a great example. Clutter all over. We do not have the workforce. We know that there are so many backlogs and that is one of their biggest challenges. So one of the things I think we could be doing is using our procurement authorities and our Grand Challenge authorities to dramatically accelerate contact rich robotics. And I think it will take some time. Obviously this is not something that will lead to a self driving car the right tomorrow or a robot that will finish the job tomorrow. But grand challenges that are designed the right way have a way of being a forcing function. The DARPA Grand Challenge did not produce autonomous vehicles by itself, but it did produce the. The workforce that went on to create BAMO and whatnot. It produced a lot of technological solutions too.
- [28:06]
B
I think people might not appreciate how there's a deep history here, right? I mean, you think about semiconductors. A lot of early semiconductor firms, they got contracts from the military because the military wanted to make satellites or missiles. And these chips were too expensive for most civilian applications. But the military purchased them and then allowed those companies to also develop commercial applications. Early Internet was the same thing. You had the arpanet, which would have been way too expensive for any private company to build. But the Defense Department was kind of the first customer. And then they allowed it to be done in an open way. You could imagine the same thing here the military needs some robots. It's willing to pay millions of dollars per robot, but then it opens the data or allows the companies to also have a commercial civilian kind of arm.
- [28:48]
A
And that's like you're buying down the risk in that way. The military has a lot of authorities to do this. I think Ann Neuberger wrote about this and she's done really good job in an article Martin and Ann wrote for Ann Reason on the robotics race with China. And they hinted at this, but they didn't go fully deep on how important the Defense Department's role here could be. To your point on semiconductors, they were purchasing what, 95% of integrated circuits in the 60s and 70s? Yes, they needed the semiconductors, but they were also setting the specifications. They single handedly created the industry from that perspective. Same for Internet. And I think there's just a rich history here where we can further use the authorities that exist. And it's also not like we're just throwing the money away. We actually need these things for some of the most important problems the military has. And those problems aren't necessarily like robots with machine guns. Those problems are, hey, a robot that can help me weld a ship, because I have a dramatic shortage of cleared welders in this country.
- [30:12]
B
So inspection repair of vehicles is one that we've, we've talked about. Are there what are a couple other examples of like types of robotic applications that you could see the military using?
- [30:24]
A
Yeah, I think that's one. There are. The military currently does use several kinds of robots for like in disaster recoveries where it is really dangerous for a human to go, and robots that can crawl in tight spaces. That happens where the specifications are basically created by them. The entire DARPA grand challenge was a whole military procurement thing where Congress had told the army to move to autonomous
- [30:56]
B
vehicles because for like supply missions they wanted to be able to move stuff around without putting a driver at risk.
- [31:02]
A
Yeah. And drones are another example here where the military is actively using them for reconnaissance missions and a couple other things.
- [31:14]
B
Yeah, so I guess that's why you were mentioning earlier that the, in China they've got these warehouses full of robots. I mean that's maybe the, maybe the Chinese government is a little more motivated or proactive about doing some of these things. But that seems like maybe a similar kind of project over there.
- [31:29]
A
I think people, I think people do not give enough credit to them. I think the Chinese robot half marathon marathons, they look funny at times. When the robots fall, it's easy to make fun of them, but they are very interesting forms of these grand challenges. At the end of the day, you're solving for balance, you're solving for battery life. You're basically incentivizing companies to create robots, inform factors that can remain stable for long periods of time. Do not need you to put lubrication fluid every three miles or so or every mile, change batteries every 10 minutes or an hour. So that is a way the state over there is accelerating innovation. Xi Jinping is very obsessed with robots. He, he went to this football match where robots were playing football with each other and he keeps asking his people, how high can the robot jump? So I feel like we haven't seen that kind of an initiative here. We did see the first lady walk out with a figure humanoid. And I think that's good. That's interesting to see that kind of an initiative. It was a little creepy. I think we need to solve a lot of things before we can solve for the appearance. But simultaneously we have other challenges that say the PRC does not. I was looking at how many state bills are there right now that what I would characterize as human operator laws. Roughly 350 where people are proposing legislation in states requiring a human operator, whether it's behind the wheel or with a robotic arm or whatnot. Only one of those laws has a sunset. That's a law in Kentucky that will sunset in June 2026. Many other laws have been vetoed. Governor Newsom, California has vetoed like two laws already that would have required a human behind the wheel and self driving cars. Tom Steyer very recently came out and said that he would just. I don't think I'm remembering it correctly, but the whole gist was like he would either ban fuller automation or require a human operator. So the politics here is very different than the politics in China. When accidents happen and accidents will happen, the political economy will shift dramatically. Whereas a high speed rail accident in China did not change a thing. Now I'm not saying that we just discount those lives that are lost. The point is to how do we continue to work on the technology while minimizing those risks and rather than just like shutting it all down.
- [34:32]
B
So another big difference I think, I think the US it's safe to say, has some of the best software, the best AI models, But China really is very strong at hardware. We've talked in the past. I recently purchased a unitree like Robot Dog and it cost about $4,000. And here in the US we have a company called Boston Dynamics that makes a similar product, probably a Better product, but it's what, $70,000. Talk to me about the hardware difference. Where does, where does China's advantage, Am I right that they have an advantage and if so, where does that come from?
- [35:05]
A
Yes and no. Right. Different things there on where they have advantage and where they do not. I think one of the critical things that I've seen with regards to China on their supply chains, I think they are no longer just a market in hardware, they're volume setting manufacturing system. Let's say they have domestic industrial OEMs whose market share, whose domestic market share has grown, you know, to over 50%. They are localizing components. Servo localization in fact is like very strong in China. Innovense is a company that holds roughly 30% or so of the of China's general servo market. Their reducer dependence has weakened materially. In volume terms, their domestic RV reducer share is reportedly.
- [35:57]
B
Let's slow down again. So explain what a servo is and what a reducer.
- [36:00]
A
Yeah, so a servo motor like so there are several components that you really need to actuators servo motors, precision reducers like harmonic drives.
- [36:09]
B
These are all different kinds of motors that make the robot move around because
- [36:14]
A
every motor operates really, really, really fast. And for a joint you need to slow down the motion. And so a robot is just like a bunch of these motors at the end of the day. And China is trying to localize a lot of them they've built. I think the the correct frame maybe five years ago was that China depends on Japan for all of these. And the correct frame now is like China is closing fast everywhere. Say that everywhere except the hardest precision layers. Leader Drive is one of their companies that is coming up like a lot where they're producing cheap yet effective motors. To your point about unitary, they've got vertical integration throughout and most of the things and they're producing everything in house where they're able to iterate on a component so quickly that it just makes it so easy to do the final assembly in a competitive price range. I talked to a couple of people in the Bay who believe that they can bring down their product pricing to roughly 5 to $6,000 as well. If they start to do in house production, we'll see where they land. But we have at least three different models in the United States. You know, you've got Boston Dynamics to your point, like they have allied production. Hyundai is providing the motors, they are manufacturing. A lot of that figure is trying to do everything in house. And Tesla is actually contracting with Leader Drive, the Chinese company for all their motors. Most of Tesla's supply chain that has been publicly reported is going to come from China for the Optimus Humanoid. So I think China has started to grow as a player here. It's quite interesting to see where they are starting to own the choke points. Some of the choke points, they're not there there yet. Like Japan, South Korea, Germany, they still produce a lot. Harmonic Drive Group, for instance, it's one of the biggest names. They have facilities in Germany, Japan and Boston, one of their largest facilities. They're doing a lot of great work. They are facing a lot of competition from Leader Drive now.
- [38:48]
B
So one thing I've heard is just that China for the last 20 or 30 years has had the lead in consumer electronics, smartphone, more recently, EVs and that just that supply chain also I think for flying drones. DJI is by far the market leader globally and that that kind of cluster of suppliers and those tense, those dense supply chains and social networks give them just a big leg up when it comes to robots. Do you think that's right?
- [39:19]
A
Yeah, I think so. I haven't heard a good example or a good reasoning for why it's the case outside of that. It's really easy to iterate on a component in Shenzhen, where if you are an OEM based in Shenzhen and you're producing say a humanoid, you can walk up to your component supplier, get repairs done or a new iteration of a component within a day. If you're doing that in the United States, you have to like send the product back. You have to send the component back. Sometimes you have to wait weeks, wait weeks or months before you will receive it. It's just too costly, the iteration is slower, you end up being behind. And so how do you not just produce more robots in the United States, but also bring some of that component? Manufacturing here is a really important task to think of. And one of the ways this is a theory of mine that how this could work is if there was enough demand for the final product in the United States, there would be economics for the component manufacturers to come here as well. But if the demand is just for the Chinese robots, the demand remains for just the Chinese robots. Because they're cheap, they're cost efficient, then it makes less of an economic sense. And that goes back to our DoD point of how DoD could create that demand.
- [40:50]
B
Yeah, I mean, this seems to some extent just a flip side of the Silicon Valley story where you look at any particular point of the Silicon Valley ecosystem. There's nothing that special about it you've got a bunch of like investors, you got a bunch of programmers and like lots of other cities in the US and elsewhere have like tried to create innovation hubs or whatever and they just never get the same traction because once you just have, you know, like once you've got a certain critical mass of engineers and entrepreneurs and investors in one place, they can have coffee easily, they can quickly go from one firm to another and it's just really hard to dislodge that. And my sense is like Shenzhen just is the consumer electronics hardware version of that. It's just theoretically you could build Shenzha anywhere else but just there's a lot of gravity once that center is established, it's really hard to get it re established anywhere else.
- [41:39]
A
And there's a lot of cassette knowledge that comes with producing things that you never write out. Not to go full on philosophy here, but there is merit to that and we are realizing that in critical minerals now this is true of manufacturing too. When we stop producing these components or doing consumer electronics manufacturing in the United States, we lost a lot of that asset knowledge on how to iterate on things. And that will only come with more reps. It will take time.
- [42:09]
B
So something you wrote that I thought was really interesting was that it's not necessarily that the actual components are made in China. You were saying, I think you said this before that Japan and Germany, I think in particular maybe South Korea, some of these US allies that are maybe near China but not in China are actually the source of a lot of these products. But then often a Chinese assembler will buy parts in Japan and ship them to China, put them together in China and then ship them back out to
- [42:36]
A
a gold market, a lot of them do that. If you look at made in China's targets, for instance on localization of robotic components, the made in China target for localization by 2025 of say ball screws, precision actuators of all these components, CNC systems, everything was at 70%. And US China Commission did a study where last year, I think in November, it's very interesting where they evaluated China's performance on made in China and none of these robotic components actually got there. They were roughly at anywhere from 10 to 30% of localization. But they are accelerating that domestic localization, but they're not there yet. And so they're starting to build it, but they still continue to import a lot from Japan, South Korea. They're making progress where it matters most for their domestic market. I think leader drive, for instance, what I was Talking about their share of China's harmonic drive market, I think like roughly tripled in the last eight years or so. And it may have even crossed harmonic drive, like declining China share. There are a couple others who are coming up, like Shenzhen, Hans Motion Technologies, another company that is an emerging player that is also producing stuff. So this matters for Chinese humanoid production costs a lot. If they can produce more of these components there. I think you bought a unitary dog today for $6,000. It's possible. If a lot of this localization happens and they are able to bring down the costs even more, you may see the dog's price come down even further. And the other thing there is we don't necessarily account for all the indirect subsidies that are going in the system that are leading to that price. I talked about the 40 factories that China had announced and two dozen of which are operational. The state provides the land, the factory floor massively subsidizes the workers. Compared that to say purchasing data from an American data collection company. Very different. The state owns roughly around 10% of Unitree as well. And so there is a lot of state involvement in the development in the economics of these companies in China. And I'm not saying we should try to copy them. We should definitely not try to copy them. It's not the right playbook. But they are running a playbook and we need an answer. We don't have one. Everybody keeps talking about, hey, we need a national robotic strategy. I meant so many lunches and breakfast where somebody has altered that phrase and then has no follow up to explain what it means. It's like the classic Washington version of like let's create a commission, you know, another robotics commission.
- [45:35]
B
Yeah, let's. So, so one thing I wonder about that you're mentioning that there's a lot of, of capacity in US allies. Like I wonder to the extent that we're, you know, built, somebody's coming up with a plan for the US to become a leader in robotics. To what extent it should be trying to build up the American one versus maybe we don't care that much about the American. We just want to make sure that the Japanese and Korean and German and whoever those companies are healthy and that we knit them together into a supply chain where all the work is being done in American knowledge. I mean, I think there's a similar thing with shipbuilding, right, where it's not great that the US doesn't really have the ability to manufacture ships, but like Korea and Japan are good at manufacturing ships and maybe we should just be buying A lot more Japanese and Korean ships. How do you think about that?
- [46:19]
A
I think there is merit to a strategy where in the near term you continue to buy, but you do want to get to co production and ultimately you want to get to some level of domestic production for nothing else but the tacit knowledge, if that's the only thing you're getting out of it, because you don't know what may become a supply chain choke point tomorrow? Did anybody think that PPE would be a supply chain choke point until Covid hit where a country would then decide to weaponize those supply chains against you and be like, hey, you want our PPE or this vaccine? You gotta derecognize Taiwan, right? And now with critical minerals, which we've seen actually since 2010, happening multiple times, say the Japanese prime minister makes a comment about Taiwan and China then imposes more rare earth restrictions on them. So I do think like from a supply chain security standpoint, you want to diversify your supply chains, but you also do want to have some level of those supply chains, at least some capacity domestically. Now you won't be able to have everything domestically. We just don't have the workforce. There are physics constraints on resources and whatnot. We are struggling with the load growth that we have right now. And so like all these challenges will have to be solved, but that's not a reason not to have some of that.
- [47:48]
B
So talk to me about why this matters. I mean, you know, all our iPhones are made in China and not all of them, but many of them are made in China. And it doesn't seem to have caused any.
- [47:57]
A
All the American ones are now assembled in India. The components are coming from China.
- [48:01]
B
Okay, my mistake anyway, there's a lot of manufacturing in China. So far it doesn't seem like anything catastrophic has happened. Why is it important for the robots of the next 10 or 20 years to not pretend and not to end up as the dominant provider the way they do with drones and a lot of consumer electronics?
- [48:17]
A
Yeah, that's a very good point. And I think the answer to that lies. And I'll tell you a story of last year, April, there were two security researchers. They basically were investigating a unitary robot dog. One of those that you have now. It's the go one. It's a quadruped. That's like roughly the size of a beagle or something. Like one of those small dogs.
- [48:45]
B
Sounds about right. That's about how big mine is.
- [48:47]
A
Yeah. A lot of university labs have. And if you look at actually Unitree's IPO they talk about, oh, one of our biggest purchasers is a major University founded in 1900, major research University founded in 1900 in the United States. And I'm like, well there's only one that's Carnegie Mellon major research university founded in this specific year. Whenever Cal was founded. You can tell University of California, Berkeley, Carnegie Mellon, Cornell, NYU are some of the major purchasers of unitary robots. And what happened when these two security researchers were investigating this GO one quadruped. And by the way, actually a lot of law enforcement are using it too. They identify a tunnel service that was pre installed that auto activates on Internet connection. It is invisible to the owner. So anyone with an API key could watch through the cameras. Anyone with the API key could take control of the robot. There were roughly 2000 devices, something like that, like 1900, 2000 devices that were exposed at universities and prisons defense installations. And that was not the only thing. Like six months later there was another team that was looking at the newer robot, that's the GO2, the B2 as well. And then the G1 humanoid. They found that one compromised robot was able to scan for every other unitary robot in the Bluetooth range and infect them automatically. These were all exploitable security failures. But then there was a third finding that same year, like last year itself, in the fall, autumn, fall, whatever. Like There was a third team that was looking at G1 humanoids, what they do when operating normally without any exploit at all. So the robot was transmitting data to servers in China at regular intervals during normal operation. Audio from the microphones, video from the cameras, GPS coordinates, all the torque readings, spatial mappings from the lidar, everything like. And if the connection dropped, it was reconnecting automatically. There was, nobody was being notified. There was no consent mechanism. People were, there was a firmware update afterwards. The researchers tried to see whether that resolved it. It did not. So there are things that are in these robots and I think it's one of the, the things that is important. Background here is the national intelligence law of 2017, the Chinese National Intelligence Law that compels every individual and organization operating in China to comply with Chinese intelligence activities. Everything here is going to the MSS at the end of the day. And unitary is not an outlier. These companies have, they have been acquiring a lot of access to our spaces through products that we're all buying voluntarily connecting into our own networks, whether it's TP link routers that FCC just took an action on or DJI drones that it took a long time to ban. And speaking of dji. Actually, DJI also has a robot vacuum. It's called Romo. This February, a researcher was able to access what, 7,000 units or something in roughly two dozen countries. Whether it was two dozen or 25 countries, I don't exactly remember. But he was able to access all of those microphones, all of the real time floor plans. There was no hacking required. And so the national security case is very real here. The Article 7 of the National Intelligence Law and all that. And then you've got state ownership of a lot of these companies. A lot of these companies are also like, you know, sitting on China's National Humanoid Robot Standardization Technical Committee, which China's Ministry of Industry and Information Technology announced last year, I think in November. That body is supposed to be writing the rules for Chinese robotics development. Now it includes execs from Unitree, Yubytech and also Huawei ZTE SenseTime, Huawei ZTE SenseTime. All three of them are on US entity list. China Mobile, which is designated by the United States as a Chinese military company, is also on there. And there are also like defense industrial research institutes, PLA linked universities. So the boundary between the commercial robotics and state security apparatus in China is basically by design. It's porous. So when we are talking about whether it's okay or not for us to be dependent on Chinese robots, I think the national security case is really strong there. That this is just. You take the example of the connected vehicles rule that Commerce Department published a while ago where it banned connected vehicles from China coming into the country. And the rationale there was simple. Like all of them connect to Chinese servers like the Mothership. And the national security case is real with mass surveillance. And also like it's a two ton, you know, block off aluminum steel that's just idle on the road. If the the Chinese government decides to stop all these cars one day with robot, the action becomes thousands of these across everybody's homes that can act, you apply force. The case is actually much stronger. And yet you would think that we would have taken an action on PRC robots, but we haven't. There's no ICTS rule, there's no covered list entry, no CFIUS action either. Earlier this year, for instance, Shenzhen Paisia acquired IROBOT out of bankruptcy. Shenzhen Paisia was iRobot's biggest contract manufacturer. They acquired it out of bankruptcy. There was no CFIUS action. Every American's home data, every person who had iRobot, well, your home data is in the hands of the Chinese now. Congratulations. The US government did Nothing.
- [55:14]
B
Yeah, I think this is a very legitimate concern. Some people might underestimate the intelligence value of even pretty mundane details about people's lives. I mean, most people, who cares? But if you have people who work in national security type roles or you have somebody the Chinese government decides to want to spy on, it's very useful to know what's the layout of their house, what's their day to day schedule. People have affairs. Maybe some robot sees you having an affair, that's like a blackmail opportunity. It's just very useful when you're doing intelligence, like little bits of information and even seemingly innocuous bits of information. When you have a lot of it, you can put it together and learn things that are actually very useful for various kinds of intelligence gathering.
- [55:53]
A
And I mean, we do massive mass data gathering on foreign nationals. We know this works. That's why we are concerned.
- [56:05]
B
You don't even have to be that paranoid about the Chinese. You don't have to think that badly. The Chinese government, I mean, the Snowden disclosures, like the NSA used the fact that US Internet cables run through the United States and that all these tech companies, like sometimes the NSA compels American tech companies to turn over information about foreigners or to surveil as foreigners. And as you were saying, there's even less. I mean, there's some due process at least with that in the American context. In China, that's the way the system works over there. The Chinese companies pretty much have to do what the, what the government tells them to. And certainly we're not going to have any transparency into when and how that information is used. So yeah, I think this is a very legitimate thing. And then the other, I guess is probably obvious. But also if we ever did get into a conflict with China, if they're able to produce 100 times more drones than we are, that'd be very bad
- [56:49]
A
for that would be really bad. Or the fact that say if the US army was, or the Navy was only purchasing unitary robots or something and they suddenly stopped working in the event of a conflict, that would be pretty bad.
- [57:04]
B
What's this look like internationally? So it does seem like the U.S. i mean, the U.S. has limited certain kinds of Chinese products like cars and hopefully more in the future. But there are many countries where American and Chinese products compete against each other. Are there other countries that are taking the kind of stance the US has that we don't want Chinese stuff here or most countries like happily adopting Chinese technology in these areas?
- [57:29]
A
I think it depends on what kind of technology we're talking about, right? Like Europe still very publicly uses Huawei. And it's as much as their own companies have been telling them that this is bad. We know there's Huawei in European networks. India was the first one to ban several Chinese apps. Banning TikTok way before the US even decided was a problem. And they've now recently banned all Chinese CCTV cameras because they had realized that the Chinese use, like, CCTV cameras to plant malware in the Indian grid. So they just were like, okay, we're going to ban Chinese CCTVs. And I think a lot of countries are going to start realizing this more and more. But there are some areas where it's just like, really hard to do something. We've realized that it's really hard to do rip and replace. It's easier to ban something before it becomes widespread adopted technology. Connected vehicles rules successfully went into effect because there was not a lobby that was there to lobby against it. It was not the local police departments that were lobbying against it. DJI drones were so hard to ban because local police departments would come to the Hill and be like, no, I love my DJI drone that was gifted to me by the Chinese. Yeah, sure, it was gifted to you by the Chinese because they love the data you're sending them. So I think that is a key theme that people are starting to realize that we need to get ahead of the curve rather than follow the curve. Identify the technologies that are going to be a problem and take an action on them ahead of time. I think a lot of countries are starting to think of that. What does it look like in trusted tech space, trusted access broadly, and countries that, interestingly to me, countries that share borders with China are more inclined to take those actions than countries that do not share a border with China, because those countries have actually realized what happens when you share a border with China, whether it's their military constantly encroaching in your naval domain or across the border or whatnot. Whereas people here do not necessarily see that or feel that. And so it's just like, no, what's wrong with my TikTok? You know, I love my TikTok. I love my Red Note. I want my Xiaomi or like, I want all the slave labor Temu Shine fast fashion. And my smart display is coming from China, where the microphone always stays on and it connects to every device in your home. By the way, like 90. What? Like 80, 90% of displays come from China. Like, so there are a lot of dependencies. And now I'm not saying the answer is to boil the ocean that go after everything. But you do have to think of like, what are the non inert products? Where. Yes, okay, fine. Like screws and bolsters coming from China. Those are inert products. We can tackle them later. Toys, furniture, I don't care, we can tackle it later. But like the smart stuff, the connected devices, anything that's sending a radio signal, let's think about how we can diversify away. Optical transceivers. I was just talking about on a panel earlier about energy supply chains. Majority of the optical transceivers come from China. There's one American company that produces them. And so let's think about it. Whether it is okay for us to be incorporating this one critical component in the electric grid and in our GPUs, this one component being from just China. Let's think about all those things.
- [61:43]
B
Do optical components have the ability to transmit data?
- [61:47]
A
Yeah, so they send a lot of null packets along with like. So when they convert electrical signals to optical signals, they also add a lot of null packets. And you could replace a null packet with actual information and you would not see anything, you wouldn't notice anything unless you were investigating the packets themselves.
- [62:06]
B
Because it seems like a good design line is like, has a network connection or doesn't have a network connection. Because if it has a network connection and especially if it has the ability to do remote software updates, which most like modern devices have, then even if you like, like sometimes I'll, you know, I'll be talking to somebody about this and they'll say, well, you could like audit the device and see and like one that's like really hard to do, we're not going to do that for a bunch of consumer electronics. But also if there's a software update ability, even if you audit it very carefully and there's nothing suspicious right now, if some Chinese company has the ability to push new software on, then like you, you just don't know. Like there's no, it's just impossible to beat.
- [62:42]
A
One way to think about this is like, even if there are no security vulnerabilities right now to your point, right, let's say you have a warehouse robot. That's where the software updates are coming from China tomorrow. I suddenly push a software update that reduces the or throttles the robot, say 10%, your output in your factory reduces by 10%, your packaging, 10% fewer packages and whatnot. You didn't notice anything. You don't know. You can't investigate that. You just saw a software update coming through. And you said, yes, I'm. In the time of war, there are 500 critical factories who are, say, producing these critical components for the US Military. Operating Chinese warehouse robots all receive software updates throttling their production. That is a bad thing. That may not be evident when you're purchasing it, but it is a possibility you have to guard against.
- [63:49]
B
And this is the point you actually made in one of the pieces that I appreciated, which is that this is a change from the way robots used to work in the past. You had classically designed robots, had explicit algorithms that said, if this, then do that. But now you've got these deep learning models where nobody in some sense knows what's happening under the hood. And maybe if you had infinite resources, you could audit them somehow, you could test them somehow and figure out what they're doing. But in practice, like, there's no simple way to look at it and say, does it have some kind of backdoor for the Chinese government, or is it like doing something you don't expect? You just have no idea. And so that means, like, the, the provenance matters more. You want to have these opaque models coming from an organization you trust that you think is going to be loyal to the US or to a US ally. And outsourcing that stuff to the Chinese does not seem like a great.
- [64:34]
A
Yeah, exactly. And to your point. Right. This is really important to think of in terms of trusted tech broadly. There's a reason why we prefer Nokia and Ericsson in our telecom because we know the Chinese keep hacking it, keep hacking into our telecom systems, just as
- [64:54]
B
we're trying to hack into their telecom system.
- [64:55]
A
Yeah, and we know that because we are trying it too, just as the President said. I didn't say that the President said it. But these are really important questions, hard questions to think through. There are not a lot of clean answers to a lot of these, but hopefully we'll get there one day.
- [65:13]
B
That's it for this episode of the AI Summer podcast, which is edited by Vulgate Media. You can find more of my work at my newsletter@understandingai.org support. I'll be back soon with more voices from the AI summer.