FCC robot ruling shines a spotlight on U.S. policy; how next-gen AI can help warehousing

0:00Every growing Postgres database eventually hits a wall. Queries slow down, dashboards lag, and teams consider adding a second database. Tiger Data, creators of TimescaleDB, extends Postgres with time series primitives, columnar storage, and automatic partitioning so your queries stay fast on live data. No pipelines, no migration, no second system. Just Postgres, built for the workload you actually have. Try it for free at tigerdata.com slash go slash trial. What's up, everybody? Welcome in episode 255 of the Robot Report podcast. I'm Stu Crow, executive editor of the Robot Report. Thank you, as always, for tuning into the show.

0:42Hope everybody's having a great week. Special fisticuffs edition of the Robot Report podcast. Joined by Mike Oitzman, senior editor of the Robotics Group. Also with us is Gene Dimitri, Editorial Director of the Robotics Group. Gentlemen, what's going on? Hey, welcome. I go away for a couple days and come back and the world's on fire. The world is on fire. Gene, how goes it? It's good to be here with both of you. Yeah, it's good. We have a great show. We're going to get into, there's really only one news story of the week, right? It's these, shall we call it restrictions? Gene and I, you and I were going back and forth on this yesterday when we published this story, the restrictions that the FCC has now implemented on foreign advanced robots.

1:28So we're going to dive into this here today. We're trying to learn a lot about what's going on. There seems to be a ton of confusion around this. So we're going to continue to dig into this issue, not just here on the podcast, but we'll do that in a little bit. And we're also going to play an interview later on with Derek Pridmore. He's the CEO and co-founder of Osaro. So we'll play that a little bit later on, but Mike, tell folks what they're going to hear in that interview with Derek. Yeah, this is a great conversation Gene and I had with Derek a couple of weeks back. Again, he's the co-founder and CEO of Asaro, so deals a lot with the business that Asaro is doing with their robots.

2:04And he's going to break down for us how warehouse robotics has evolved from the limited perception systems of the past now to adaptable AI-driven automation. They've been And through all of this, we talk about holy grail applications that, you know, they've solved in the past. It was bin picking each is singulation, all that sort of stuff. And where the next frontier is on the AI side with AI driven robotics in the warehouse. So it's a great and fun and educational conversation with Derek. Orsauro has been around for a number of years now. So looking forward to hearing that a little bit later on in the program.

2:41And before we dive into the news of the week, just a reminder about RoboBusiness. This is one of our flagship commercial robotics development events. It's the sister show of the Robotics Summit and Expo. RoboBusiness takes place October 20th and 21st out in Santa Clara, California. The 20th anniversary show, gentlemen. We've been involved in how many of collectively? Probably 30 of the number of times each of us has worked on this show. So we've been along for a lot of them. So this is going to be a fun one. Over 2000 attendees, over 100 exhibitors on the show floor. Again, talking all things commercial robotics development.

3:19We're going to work in some fun look backs at the 20 years of the show in the robotics industry as a whole. So it should be fun. If you check out RoboBusiness.com, you can register, you can buy a booth, you can reach out to us if you're looking to partner or get involved in any way, shape or form. We have the first 40 speakers up there. We're going to still add a bunch of people over the next few days and weeks, but make sure you check out an amazing lineup of speakers so far. Rodney Brooks, Ken Goldberg, Jonathan Hurst. We have Universal Robots. We have Amazon. We have Teradyne. Aptronic is going to be there.

3:54Morgan Pope from Familiar Machines and Magic, the new startup, he's going to be there. So a lot of amazing people are going to be there talking humanoids and physical AI. And I think, again, gentlemen, chime in here. But what we're trying to do is really these are the best of the best. These are amazing people, but also amazing companies who have been down this road. They have all the scars from how difficult it is to build reliable robots that can scale to a significant size in different applications. So we're trying to learn from them. Right. How is Amazon using AI? right now in their robotic system.

4:32So that's what we're trying to do, get the best of the best together to talk about all things commercial robotics development. So again, we hope you'll join us at RoboBusiness October 20th and 21st out in Santa Clara. Reach out to anyone here at the Robot Report and we'll help get you the answers that you're looking for, but we hope to see you there. All right, on to the news of the week. And again, we have you folks covered with our network of websites. Make sure you check us out at the Robot Report and Automated Warehouse. Follow us on all the socials. Sign up for our weekly newsletters. There's no better way to stay up to date with what's going on than doing that.

5:05As I said off the top here, really only one story that we're going to focus on. And Gene, I think I'll let you have the floor on this one since you reported on this for us. And you got a bunch of commentary from different experts around the industry. But the FCC is limiting U.S. imports of advanced robotics from any foreign country. and there was a lot of people, including myself, who misread this at the beginning, we'll be completely honest. I think the overarching message that you'll see in the media is that this is limited to humanoids and quadrupeds. That's not true. It's literally seems like, I mean, drones are already banned, right?

5:43It's any sort of robot that can move around. So it could be a humanoid. It could be a quadruped. It includes tracked robots. It includes wheeled AMRs and AGVs. So this is a lot broader than I think a lot of people understand. I think anything that has what camera and perception on it was part of the defining aspect here, right?

6:03Derik Pridmore: Well, and it's partially connectivity, right? The FCC regulates spectrum and they're coming at it from the angle of, oh, if it communicates wirelessly with anything else, we have jurisdiction over it. And so, yeah, you know, we weren't careful in crafting our headline on this team, as you mentioned. Is it just a ban? Is it just on Chinese humanoids? No. The way it's worded right now, it could be broader than that. Although how it's enforced or how the administration and how the industry interpret what the FCC is now ruling is still up for debate. And, you know, this has created a lot of fear, uncertainty, and doubt among people that we've talked to directly, among commenters on LinkedIn, across the robotics ecosystem.

6:52Derik Pridmore: We've already reached out to mobile robot providers and big robotics clusters and associations, some of whom are willing to go on the record with us, as you might already have seen, and we are going to continue our coverage. But one of the things that is behind this is the federal government claims that cybersecurity and protecting our supply chain or our manufacturing ecosystem are the primary drivers behind this. And there are absolutely legitimate concerns. We know that a Unitree robot was found to have some pretty serious vulnerabilities. and there's certainly the notion that, oh, we want to get ahead of competition.

7:38Derik Pridmore: The U.S. lost its domestic drone manufacturing capability years ago on the consumer side and there's a fear, oh, if humanoid robots or mobile robots are the next big thing, and in this I would include mobile manipulators, which might be semi-humanoid, that we need to somehow guarantee or protect American industry. But at the same time, there are some very notable exceptions to this. So mostly allied countries can apply for waivers, right? So if you're a foreign company and you make a robot that might be affected by this FCC rule, You can apply for a waiver and you're back in business. There's also the U.S.

8:30Derik Pridmore: government. The federal government is exempt. So if it's such a national security concern, then why would the government be able to say, oh, we're just going to buy whatever robot we want? To me, it's somewhat of a contradiction. And then finally, you know, Steve, I know you want to get into this. isn't really aimed at one country in particular, but it could affect imports from all kinds of companies because let's face it, all of the major industrial automation providers are not headquartered in the US. They have all expanded manufacturing in the US recently and this will probably encourage that.

9:10Derik Pridmore: And that's a good thing. But we have to be realistic about where are certain types of robots produced? More importantly, where are the component parts produced? So even the American humanoid robotics companies, which are arguably on the cutting edge of the technology, how many of them rely on foreign-made components still? So there's a lot to unravel here. And one key thing, and I don't think you mentioned it there. Maybe you did. You said a lot there, which is spot on. This only applies to new models of robots. Right. Right. Let's just stick with the unit tree example. This is the, I think, example that most people are running with, because to your point, this is not in my in my mind.

9:55Right. Again, having talked with people yesterday and having my own thoughts on this, this is not about it. Is cybersecurity with robotics an issue? Of course. Right. Do we need to strengthen our own supply chain and build up our own national robotics strategy? Of course. Right. But this is 100 percent about China. Right. Let's not. And I don't want to speak for you guys. This is my own personal opinion. This is only about China. And if anybody tells you otherwise, I don't know. But it only applies to new robots. Right. But let's stick with the Unitree example. Right. So they're they're G. I think it's the G1 humanoid is the robot that you see at all the trade shows.

10:37It's probably the go to development platform for folks at universities and other research labs that are testing AI algorithms. Think about all these new whiz bang AI companies that are raising half a billion dollars in funding. If they're if they're testing on a humanoid robot, guess which humanoid robot they're testing on the unit tree. Right. So those G1s are approved. Right. They're already here and they're going to be continued to be allowed to be sent here. Right. Which to your point, Gene, is if like, how is that possible? So if if unit tree comes out with new chemoids, which they will in the future, it's going to potentially cause problems for them.

11:18But they'll need a waiver. I mean, I think the part of the exception here is you can always get an exception, but that's the way they'll manage that.

11:25Derik Pridmore: They could, but then if they get a waiver and they're not secure, does that kind of defeat the point of having? Well, that's part of the waiver process should be an evaluation of the requirements for that, right? Yeah. So, but I think that that's, again, all of this stuff is so contradictory, right? If we're trying to rebuild our own internal supply chain as a country, we need robotics to do that, right? Do you guys agree with that? And to your point, Gene, there is not a single, again, like what's the largest? Amazon Robotics is probably the leading U.S.-based, born and bred in the United States of America robotics company, right?

12:08And they don't sell their robots externally. They only use them internally. So if you take them off the table, what's the next biggest U.S.-born, again, been here since the beginning, robotics company? I don't know the answer to that question, right? Locus maybe on the AMR front, but Universal Robots, they're not – and this is where the technical weeds – we've reached out to Teradine. We're trying to just learn – again, just trying to learn about this, right? Universal Robots has manufacturing capability in the U.S. now. Does that mean they're a U.S. company? Does that little detail get them – and again, they would be approved anyway, I would assume, right?

12:49But there's so many – I would hope so.

12:51Derik Pridmore: I certainly would hope so. Yeah. But they're owned by Teradime, which is a U.S.-based company, but they're based in Denmark, which has had its own ups and downs with diplomatic relations with the U.S. And so, you know, it does, because it's such a broad policy, it's understandable that there's a lot of confusion about it because it could apply to wheel autonomous mobile robots as well as to next generation humanoids and quadrupeds. And the other question is, we know it's not the only policy that's working its way through Washington. We know that there are attempts to try to ban outright federal use of foreign-provided humanoids.

13:40Derik Pridmore: Certainly there were concerns about drones before. The other question that I have is, are these measures the right way of stimulating U.S. manufacturing? We know it takes a while to spin up factories. We know it takes a while to make sure that you're sourcing all your parts and materials correctly. And everybody I've talked to is in favor of that. So that's not a controversial issue. But how do we go about it? And how do we make sure that we're picking winners and losers in this that were not stifling innovation. A couple of people brought that up. A couple of people brought up, you know, with the restoration of tariffs, is this one more disincentive for even our allies to trade with us?

14:27Derik Pridmore: There's a lot of challenges in the robotics industry, which has always been multidisciplinary, which has also always been a global thing. Now, yes, we do not want to be overly dependent on China. China has conducted things that were not ethical in terms of stealing IP or in terms of underpricing the market in order to just dominate certain areas. And we know that the president is planning a summit with Xi Jinping, so this will no doubt come up. But the fact that China has already threatened to retaliate, what does that do to U.S. manufacturing in the meantime? You know, we're not going to be independent overnight.

15:13Derik Pridmore: That's not how things work, unfortunately. So how do we manage that? And, you know, increasingly, when I've talked to some of our allies in Europe or in Asia, they're looking elsewhere. The US is still a very appealing market. Everyone still wants to come here, but they're hedging their bets now. And that could be counterproductive in the long run. I think that when we talk about how do we restore, you know, the capability of the U.S. to deliver and build high-tech products, there's more to it than just magnets and other things. Although, you know, the raw materials are also a big deal here. But at the end of all of this, for everybody that's listening, that's building something that's complex, you know, you've got supply chain around the world for all the various components, cameras and motors and wires and what have you, to bring together, you know, a complex electromechanical system.

16:19So, you know, at the end of the day, it's a challenge for all of our listeners to try and figure this out. You know, again, we're trying to understand here what's going to be the long term impact. Is this just a bargaining chip, as Gene sort of hinted at, you know, for the global trade interaction between the U.S. and some of its other partners? or is there going to be a real impact as we've seen play out on the drone side, right? So I think the drone ban is probably the best example. And this is, you know, the drone ban on steroids now. They've learned from that. The industry has learned from that.

17:00Many companies have pivoted in the drone space to react to the drone ban and have found success. And there have been some failures. But along the way, the industry has figured that out. I think that's what we have to look at is where this potentially could end up. Yeah, again, so I do think it's all about bargaining with China, right? And trying to force them into better financial trade situations or what have you, right? I think that's all this is about. But I also think, you know, maybe, again, this was in a discussion that I had yesterday with somebody, is how much does the government actually understand about robotics?

17:39And how far along we actually are, right? We talk about this all the time. Again, I think this issue, and again, I'm speaking for myself here, right? I think this has come up a lot. I think the humanoids and the AI stuff from China is really the cause of this. And everybody else is a byproduct of the regulations, right? But I do think the humanoid race, the AI race is what sparked this. And there's a lot of baggage that everybody else is having to deal with now. But do they understand the actual capabilities of humanoids in the U.S. today and AI companies today? I mean, again, somebody I talked to yesterday was saying that they think if you hear all these people that are just overhyping the technical capabilities and performance of both humanoids and AI, if you put them both together, there's been so much talk about this is going to solve manufacturing and this is going to solve me.

18:40And that's just not the case. Right. Right. So how much have they been? Has the government been misled about how far along we actually are? Because I just don't think this makes any sense at all to just blanket ban the sale of any robot from any nation and company on the planet. And I know there will be exemptions from ally countries and even from some non-ally countries. Right. It just doesn't make any sense, because if you're trying to rebuild, again, the supply chain, we might not want the Chinese robots working here, right? There's probably higher performance industrial arms and mobile robots that are built in Europe.

19:21Or, you know, again, Locus is a great example of a US-based AMR company, right? But where do all the parts that go into the robots come from? So much of it comes from China, right? And there's a rule. I think, Gene, did you say it again? Again, if you want to be considered a domestic manufacturer, not only are you manufacturing your robots here in the United States, but 65 % of your domestic component supply chain has to be from here. And does any robotics company meet that threshold? Again, it's the same problem that drone manufacturers had, right? Because China had cornered the drone motor market for drones, and that was a primary component.

20:06you didn't have motors, your drone's not getting off the ground, right?

20:10Derik Pridmore: Well, and that also brings up the point. It's like, again, we all agree that we want American manufacturing to be competitive. Yeah. But the question is, is this the best means? And one of the things that I think this brings up is the need that we've seen and I've reported on, we want to have a national robotic strategy. And is the FCC attempting to ban or regulate next generation or advanced robotic systems? Is that the best way of doing it or is it a point solution? Or is it an opening salvo? That we don't really know yet. But the reason that China has been able to outmaneuver the West so much of the time is that it has economies of scale.

20:55Derik Pridmore: It has centralized. They have their five and 10-year plans. They have a centralized economy. and they can say, we are going to take the lead in robot manufacturing and we will build factories. And even if they're going to take a loss, we're going to do this to capture the global market. And the US cannot and arguably should not do that. But then how do we compete with that? And is banning one type or class of robot the best way of going about that? Or will it simply drive Chinese robots into everyone else's market? And then because we have said, we're going to rebuild our capability. In that time that we're trying to do that, everyone else gets the lead.

21:37Derik Pridmore: That's the same question that we're seeing with AI as well. Yes, we have legitimate security concerns about Chinese AI. We have legitimate economic and competitive concerns about it. But if we shut ourselves off from it, we don't necessarily know what they're doing and we're not necessarily competing. We're just competing in our own domestic market. And so there is the concern about how do we maintain access, how do we encourage people to come to our market, and then how do we make sure that we're keeping up with innovation and competing on the level playing field. And this, to me, says a national strategy that involves all of the agencies of the federal government and that has a unified, consistent communication.

22:24Derik Pridmore: Okay, we're going to restrict here. We're going to open things up here. We're going to do, whether it's tax stimulus or outright funding, which we see through the Defense Department of War. There's a lot of different levers that the government has and can use, short of a centralized economic policy. That could help the robotics industry and could help users, more importantly, of robotics. How do we encourage manufacturers to adopt these next generation systems and to have a say in what they actually need? right we protected our automotive and aerospace industries because they're viewed as essential i think you can make an argument that robotics is an essential industry for the next generation but how do you go about that and how do we get the most bang for our buck that we don't raise prices because in the short term that would seem to be a possible effect is that might raise prices for automation in the u.s and again china continues to undercut us as far as the price of its humanoid.

23:24They may not be as capable, they may not be as safe, but to a researcher anywhere in the

23:29Derik Pridmore: world, those are not as high priorities. And so, you know, we have to think about how do we better coordinate so that the left hand knows what the right hand is doing in the government. That's a great, you know, a couple of points. We had leading robotics companies reach out to us yesterday, Gene, right? Asking, hey, this doesn't include this type of robot, right? And we responded like, no, actually, actually it does. Right. You know, and I think I highlighted the exact line and form factor for this company's or one of their robots. Right. And so I think it's a mess. Right. Did people expect this to happen?

24:11I don't think people in the robotics industry expected this to happen. Maybe folks who work in tech policy saw this coming probably or perhaps. Right. But there's so many issues. I just don't buy that this is about cybersecurity. There's been many instances in the past of very reliable robotics companies where, not Chinese robotics companies, European, American, where their robots have been hacked and exposed. right it's not just robots from China like there's all sorts of other issues around security for these things um so there's better ways to deal with that and to rebuild the supply chain there's so much talk about Gene you and I talked I was on my soapbox yesterday talking about this with you but like nobody even wants to work in manufacturing in this country think about how long we've talked about that for right nobody wants to work you know are my three kids growing up looking to work in a manufacturing plant, whether it has robots or not.

25:17Like there's so many other way more complicated and fundamental issues to rebuilding the manufacturing and supply chain in this country that I wish we would start there, you know, or, you know, you mentioned like KUKA. KUKA is viewed as rightfully so a leading robotics company, right? They're a Chinese company. Right. They're owned by Medea. This happened years ago. They're a Chinese company, right? They're based in Germany, but they're a Chinese company, right? So think about ABB. They're for sale. They're, I think, being acquired by SoftBank. It hasn't closed yet. Like, why didn't we buy ABB, right?

26:00Have some sort of plan that incentivizes or just brings a robotics company here. And, hey, I know they didn't start. You know, ABB was not founded in the U.S., obviously, but bang, we now own them and they are going to be our go-to robotics company for all things U.S. manufacturing, right? And it could be somebody else. Incentivize Teradyne and them, right? Do things that incentivize U.S.-based companies to become – and maybe that's part of this, right? I will say, to be fair, the big, the top four or five or six robotics companies in the world, even though none of them is based here, they are expanding manufacturing.

26:44Derik Pridmore: I linked it in my story, and there have been a couple of announcements since then where Teradyne and FANUC and Yaskawa and Mitsubishi and all of the others are expanding their U.S. manufacturing. So that is good news. But in the meantime, that's just one step, right? It's how do you make it so that all kinds of manufacturers are able to benefit from this? And how do you make it that a supply chain? At the airport, you can assemble a robot or a car or an airplane in America with the help of robotics. But where are you getting those parts from? Can we make those parts? And for a long time, really high performance and affordable parts have not been made here for decades now.

27:31Derik Pridmore: And, you know, certain policymakers might have been looking the other way when it happened, or we might not have realized what we were losing at that time in the name of globalism. but to try to bring it back is going to take a concerted effort and it also has to be done in a way that is fair to our allies and that is sustainable over the long term longer than the election cycle because okay the FCC has announced this limitation but it's a midterm election year will it survive the midterm elections? will it survive the next presidential election? The way the economy develops should not be completely independent on who happens to be sitting in the Oval Office.

28:17Derik Pridmore: We have to have a more consistent and broader policy that people can actually depend on. And that's where, going back to the point of the national robotics strategy, I know Congress has not yet approved that there's likely going to be some sort of commission to study it for a year and then make recommendations and then make policy. I forgive me, but I'm a little bit impatient having heard this for several years now. I want us to fight the bullet and say, OK, if we know that AI and robots are the future, let's go all in. Yeah. And I feel like this is a bullying tactic. Right. And again, I'm not speaking for you guys.

28:55This is just me. But it's, you know, we're obviously losing the robotics race to these other countries that we've talked about. Right. And that was a that was a decision that U.S. companies made years ago. Well, before any of us three were covering robotics, they were a strong US-based robotics company. And there still are. But those decisions were made to outsource this stuff to other countries. I had an interesting conversation with somebody yesterday. And this is my final point, I think. But to your point, Gene, let's maximize this. What about incentivizing hardware manufacturers? Think about all these companies right now that are raising money, that the investors are just pouring hundreds of billions of dollars into its software companies, right?

29:39So all the talk is, and it is hard to build hardware. It's so hard, but all this funding, everybody wants to develop an open AI or a clone, right? All anthropic, all these horizontal software tools that are going to make not a couple hundred million dollars. They're trying to become billionaires and just uber rich. And what's the incentive? What's the structure? What's the path to someone who wants to build a piece of hardware, a robot? What's the path for them? It's so incredibly difficult, right? Because so many investors are just dumping their money into software, right? So that's also part of the conversation.

30:22So I think, how do you maximize the talent and the know-how and the skills to bring all this stuff back? There's so many other complicated issues that are going on to do this. I don't think it helps. I don't think it's the right. Now, again, hopefully it's the beginning. Hopefully there's other things that are coming down the road to augment this or that are part of this strategy. We'll see. But again, I think it's just been chaotic. Yeah. And Steve, I think I got a good piece of homework for everybody that's listening this weekend. And it's one of my favorite YouTube channels is Smarter Every Day.

31:00The host is Dustin Sandlin, an engineer. Great content. You know, he's a redneck from down south who likes to build things. And his latest episode is a great one. It's called The Problem with Capitalism. And, you know, he's tried to build a product made entirely in the U.S. over the last year, a barbecue scrubber. And all the issues he's gone through trying to build all the components, source all the components in the U.S. and build it in the U.S. It's a great, I think, sentient story for what's going on here about, you know, the difficulty building a barbecue scrubber compared to a complex robot system.

31:40night and day, but at the same time, if we can't build a barbecue scrubber entirely in the U.S., how are we going to rebuild our manufacturing facilities for our infrastructure for all the rest of the things that people imagine they want to manufacture in the U.S.? It's a great story. So go find that. Go watch that this weekend when you got 30 minutes. I think that you'll get deeply inside what he exposes what he thinks is wrong with manufacturing today and what we need to do to fix it.

32:11Derik Pridmore: I'll try to end on a positive note because I think part of it is we have to, like you were saying Steve, take this with a grain of salt. If this is more of a bargaining chip, if this is more of a geopolitical ploy rather than an outright ban, the odds are that the majority of robots, most of the robots that are in operation today, are not even covered. We're talking about new systems that are being developed, humanoids, mobile robots, and the like, maybe the waiver process will allow us to be a little more deliberate in how we're screening what technologies are coming in and what technologies we're encouraging.

32:51Derik Pridmore: And hopefully, it is the beginning of a patchwork of legislation and executive policy that will eventually be woven into something a little bit more coherent. the good news is that for the first time in a while Congress and the White House robotics is on their radar and that's certainly progress I'm hoping that they will engage with industry stakeholders and a wide range of industry stakeholders I hope that they will remember that robotics as with any advanced technology should help the entire country and not just certain regions or not just the companies that you already work with already have contracts and that we continue to work with our allies to develop these technologies because we still want to have access to the talent, to the technology from overseas and they still want access to our market.

33:48Derik Pridmore: So it seems heavy-handed at first but maybe at least gets the conversation going it elevates it. and um you know i've even argued that with china we should be cautious when it comes to matters of security but we also have to recognize that we are still interdependent and that's not going to change overnight and that's not necessarily a bad thing uh if it's handled correctly uh the problem is we've been on the u.s side we've been asleep at the wheel and sort of gave away the farm um we can be more deliberate in what we allow to be produced overseas and what we try to do at home. And so I'm also cautious and optimistic in that I think it's a conversation that we need to have in this country.

34:35Derik Pridmore: What should our industrial policy be like? What should our approach to robotics be like? If we are going to be shorthanded in all these critical industries, especially manufacturing, automation is absolutely part of the answer. But let's not just apply it to one class or let's not just apply it to one country. That may be the best outcome from this event. We'll obviously continue to follow this story and everything else that's going on in robotics. If you have thoughts on this story and the regulations from the FCC, if you want to talk to us, if you're willing to talk to us, again, we're trying to learn about this and educate ourselves and the industry as a whole as to what's actually going on here, how robotics companies and the ecosystem in the U.S.

35:25and elsewhere could be impacted by this. Maybe it's not a big deal. Either way, your thoughts on this, we'd love to hear from you. So reach out to any one of us here, and we'd love to learn your thoughts on the story. But again, make sure you follow us at The Robot Report and Automated Warehouse. Follow us on social media channels. Sign up for our weekly newsletters so you don't miss a beat. All right, on to our featured interview. And now you're going to hear from Derek Pridmore. He's the CEO and co-founder of Osaro. Mike, what are we going to hear from the conversation? Yeah, just a reminder of Osaro for those that don't know, warehouse automation developer.

35:59They've used robots that use AI for perception and picking of complex and varied products used. You know, they do each's picking, product backing, that kind of thing with their robot arms. Been doing it for many years now. One of the leaders in the industry. And so I think you're going to enjoy the depth of conversation around the current state of AI. very relevant to what we just talked about. And you're going to hear from Derek about his thoughts of how AI is evolving and where it's useful in the warehouse. So here's Derek.

36:32Well, Gene, excited to have Derek Pridmore, co-founder and CEO of Asaro on the show today. Catch up a little bit on warehouse applications, AI and the rest. Derek, welcome to the podcast. Thank you for having me. I'm excited to be here. Yeah, we're excited to have you as well because of what you guys have been able to do at Asaro over the last decade. Pretty amazing. You've spent 10 years building resilient physical AI implementations now. And a lot has happened in the last 10 years. I guess let's start there. How has the definition of AI and robotics changed from 2015 to now for you guys? It's hard to actually keep all the context, but it's changed massively.

37:21I mean, just as a reminder, so quickly on my background, I spent, before Osaro, I was a venture capitalist and I spent that time focusing on machine learning. So I led the Series A in DeepMind, which became the core of Google's AI effort. And now, you know, it's sort of seeded a whole bunch of efforts at all of these companies and kicked off of this massive AI wave. But when you go back to started doing that in 2010, started this company in 2015, having seen all of those really early demos and models and maybe some of the dreams of what the applications could be. But having said that, in 2015, if you recall, the models that were available were super limited.

38:07So ImageNet, which most people don't talk about anymore, was this kind of toy model that really started to exceed human capabilities right around 2015. And you couldn't even put bounding boxes on things at that point. So bounding boxes, masks, these things came around 2016, 2017. And it wasn't obviously until the last few years where you get LLMs and then VLAs, which are actually, you know, robot control models that are built with the information from LLMs. And so under the hood, what's possible with the models has changed massively, massively. We always knew that out of sorrow. And so our vision, no pun intended, from the beginning was use a cheap camera that you can place anywhere on a robot.

39:00And you should, in theory, be able to control any robot. And the model itself will not be disposable, but it will always be changing. And so you need integrations to the robots. You need great models, but those models themselves are going to always be changing. And then what you really need behind it is like a deployment layer infrastructure for collecting data, telemetry, monitoring the models and constantly not just retraining that model, but also periodically upgrading them. So all to say the models have changed massively. We have had the capability to automate warehouse applications pretty robustly, though, since 2018, 2019.

39:39A lot of these newer models aren't necessarily adding a lot to specific functionality. Most of what's happened recently is geared more toward this holy grail of generality where you can essentially just talk to the robot and have it perform all of these tasks and maybe even get better at them. So we can talk about that. There's still a lot to be done there, but there's been massive changes over the last 10 years.

40:06Derik Pridmore: Well, historically, picking was one of the bigger challenges for robots to do. And I remember the Amazon picking challenge years ago, and it was a sort of milestone moment. And compared to what's possible today, it shows how far we've come in the past decade. but let's talk a little bit about Asaro's approach to perception and you know, even though you might have seen it coming ahead of everybody else it had to be built, right? And so what are some of the approaches that you took in order to solve that perception and then enable the diverse item picking that's now possible whether it's for palletizing or for bin picking and fulfillment there's still that foundational layer, so to speak, of the robot being able to understand what it's looking at, even if it hasn't been, or depending on how it's been trained, how has that changed?

41:10Well, first you asked sort of what our approach was, and I can talk about how it's changed as well. Actually, just an interesting tidbit, Osaro actually sponsored the eventual winners of the Amazon Picking Challenge, which was an academic team from Australia that we knew. They actually solved it by building sort of a specialized machine. And so the way that we're doing it now is much more, meaning the guys at Amazon weren't expecting that. They were kind of, everyone was trying to just use an arm and solve it that way. And we can do that now. But the ultimate winners, when they, I shouldn't say the winners, that was the last challenge because they had kind of achieved the benchmarks.

41:44But when we set out to build Osaro and a platform that could power picking in a lot of different applications in the warehouse, there were kind of a few anchors that we targeted in our infrastructure. One was just integrations to all the major industrial robots. So being hardware agnostic has always been important to us. We don't build robots. We don't build cameras. as we consider those just to be components that our algorithms interface with and coordinate. So that was the first step, was integrations to every major industrial robot, which we've done. We've deployed all six of the major industrial arm manufacturers' arms.

42:30We focus primarily on FANUC and ABB these days, but we can do everything. We've also deployed Yasukawa. Just as important, in fact, perhaps more important, is on the perception side, integrations to low-cost cameras. When we started and still today, for instance, for depalletizing, most people use high-end structured light cameras. And those are not only expensive, but they have drawbacks. I mean, this gets a little bit toward the sort of Elon Musk, Tesla debate around, you know, are you going to use LIDAR or can you do it with cameras? We all know, you know, we're not shooting lasers out of our eyes.

43:06So we're, you know, the kind of human examples that we're doing it with just RGB. and then there are people who say you should use structured light. For us, our opinion is actually in the self-driving space, you should do both because LIDAR can actually get you something that the human eye can't really give you. But for warehouse use cases, it was important to us to be able to do it from low-cost cameras. And so we built a bunch of integrations, but also processing layers that allowed us to interface with, record data from, and calibrate low-cost cameras in any location. So an OSARO deployment can have as many eyes as we want it to have.

43:53Standard for us is three low-cost cameras. cameras. We use an industrialized version of the RealSense made by Framos. So it's basically using synthetic stereo. And we can put as many of those on the scene as we want, calibrate them to the robot, and the data infrastructure sort of captures all that information. And this is very important because later, as you're updating your models and building them, you need to know exactly where everything was. And again, it's easy to do that, let's say, for a humanoid where you have fixed eyes, but if you can take your eyes out of your head and move them around, that gets much harder.

44:29So that framework for doing all that very precisely in collaboration with the robot was very important. And the kind of the last piece, which is maybe the most important and where most of our IP is, is in continuous monitoring of the performance of these models. So you're literally looking at every inference from every robot at the edge, as well as all the telemetry. And you're making decisions about when should the model learn, essentially, because what you learn very quickly when you're doing this scale for, say, an e-commerce customer is that SKUs change, SKUs being stock keeping units, that's the term of art for the kind of object that you're picking.

45:10They change very frequently. And not just that, but they, weirdly and this is what makes warehouse picking so interesting um and i think important for the grasping problem is that the kinds of packaging that you see are almost perversely adversarial meaning you know if you look at a colgate box of toothpaste it's got foil with rainbow and so if you were to actually hit that with a structured light sensor it just scatters everywhere um same thing, we've had people, one of the very first things Amazon tested us with was mirrors. So you go in and they just put mirrors in the, you know, in the tote and ask if you can pick those.

45:49So having models that can quickly kick off a learning process based on performance at the edge, and all of the infrastructure do that automatically was kind of the most foundational piece of our infrastructure and still powers what we do today. So I'd say those are kind of the major pillars. And then how has it changed? Well, the robots haven't changed. That's great. The cameras haven't changed. That's great. But what's changed is the models. So there are a much wider variety of models. That's not to say they're all useful. I have a little bit of a soapbox that I get on with VCs where I try to explain to them this trade-off between generality and specificity.

46:34And there's this thing in machine learning called the no free lunch theorem. And it basically means you can't be perfect at everything without learning. And so that sort of hope and dream of a model that can do everything is always going to be a dream. You're always going to have some performance and some prior, and then you go into an environment and you start to perform and you need to learn and improve. So again, our infrastructure has always supported that. And what's changed is the types of models that we're deploying. So in the beginning, you know, we were essentially targeting use cases that had maybe hundreds of SKUs.

47:12But by 2019, 2020, we could actually perform a production with essentially unlimited number, you know, certainly hundreds of thousands and millions of SKUs where all that's needed is a small amount of learning. So typically if a new SKU gets introduced, and by the way, when this happens, you tend to see a lot of them. This is another problem. So it's not just that one new object comes by. It's usually that, oh, Colgate just changed their box sort of permanently. And now all Colgate boxes look like this. And so if you have an error and you don't learn, you're actually not going to just get a free pass and the next one's easier.

47:53You're going to keep making errors. And so what we've generally done is a lot of the improvements in learning and the kinds of models have basically gone toward us deploying models that are more robust and faster to learn.

48:07Derik Pridmore: Well, and speaking of these models and being able to learn SKUs, tell us a little bit about how Osaro SiteWorks and AutoModel enable this, because this is, you know, it's one thing that talking about the approach, but then you have specific products that are actually doing these things today. So it's pretty simple, actually, and it has to be for our integrator partners that use the vision system. We have our models wrapped in a simple API so that an integrator doesn't need to know anything about what kind of model is running under the hood. They actually work with a very simple API where they simply request actions so they can request a pick point.

48:47and they also can report back whether that was successful by any metric they want. So meaning if you got a successful grasp and that's all for you, you just want to grasp it, you can call that a success and that will feed back to the model. If you grasped the object, then you moved the robot and you had the robot do a barcode scan, but that barcode scan failed. So you had a successful pick, but a failed barcode scan. You can actually call that whole trajectory a failure. And our models will actually learn, oh, it was a great pick point, but it wasn't a good pick point for doing a barcode scan.

49:31And so it can automatically learn a better pick point. And as an integrator, you don't need to know anything about that. You just need to be able to label good or bad and report back to us. So we give them a very simple API. That's what RAP site works. and then it's all backed by auto model. So auto model is sitting there. And as you say, thumbs up, thumbs up, thumbs up. It's like, great, we're not going to change the model. And then all of a sudden it starts getting errors. And after a little while, it'll kick off and train a new model and redeploy it. You don't have to click anything or enable anything that just happens.

50:03So that's roughly how it works. Thanks for putting that in the perspective of the integrators. I know they're critical to your success every day. And we don't often talk about, you know, the role that they play and the importance of, you know, creating products that support their needs, you know, for fast implementation and the rest. And I know you had a lot of things under the hood that aren't important, just make things work. That makes their life better and faster. But I do want to dig back under the hood a little bit. And maybe you can describe the evolution that you've gone through from reinforcement learning to foundation models.

50:38in that future, where it sits today, maybe, and where it needs to grow going forward from a foundational model perspective. Yeah. I mean, the simplest way to describe it for us is we've always done reinforcement learning. It's necessary. Anytime you're doing a control policy where what you do then creates actions that change the data or the environment, you need to be continuously learning. what so what's happened with foundation models is we've simply brought them into that framework and i would just characterize a foundational model or a vla or an llm as a model that's trained on a wide variety of of data and labels that um that likely and in fact almost always have nothing to do with picking so you're it's really just a model that has a lot more data and and by virtue of that some common sense and that all comes at the cost of more accurate picking initially.

51:38So you bring in a foundational model that sort of has a much broader context. It's much more flexible, but in general, its performance is much worse. And we've seen that. So we've done implementations of PyZero. We've done implementations of Groot where we take those models. We actually take our production data and fine tune them and then run them in our office and benchmark them. And the reality is both because they're not performant and because the model sizes are so large, they're not suitable for production applications right now. They're great for demos and you can do fun stuff with them so you can actually introduce new objects.

52:16And also another difference between those foundational model implementations and what we actually deploy is that foundational models can do both the perception and the planning in one model. So, you know, you can, you're essentially not getting a motion planner. And that is nice in some senses because, you know, it's reactive, meaning you could change the scene as the robot is moving and it will adjust. But it's actually, again, not super useful for production use cases because you lose guarantees on what the robot will do. And you also lose optimality. So if I need to move from point A to point B, and I know that it's at a conveyor belt and that it's always going to be from a tote that's roughly in this position to another position, I actually can use a motion planner to optimize that movement so that it moves as fast as possible.

53:16With a foundational model, that's generally not the case. And so you'll see degradations in performance. Yeah. So what's your thoughts then on when it might become real? Do you know the actual use of foundation models in production? Are we a year away? Are we five years away? What's your sort of prediction here? Definitely not five years away. I'd say there's probably someone out there trying to do it now. And almost surely what they'll find, it's like any product that you develop. You come up with some cool algorithm and it's theoretical and super nice. And then you start putting it into implementation and you realize there's all these other things that matter more and that dominate the performance.

54:04And so having said that, even we, so I should have said this before, we don't just deploy one model. So auto model and any given deployment that we go on has a suite of models. So a typical deployment for us, we'll use sort of five models for various things. For instance, one model might do object modeling from a side camera, and it's optimized to capture images really quickly and build a rough 3D model. Another model will do the actual affordance, which is the pick point. Another model will do error detection, for instance. And we kind of run these in parallel depending on what's happening. We optimize the inference.

54:48Same as a self-driving car, by the way. They have a suite of models. They'll have the model that's looking at the lane is not the same model that's determining the pose of a person on the street. And this is a standard practice because you need to be able to inspect each of these little decisions. You can't have it all coming out of one model necessarily for many reasons. Having said all that, I think where these models will be deployed first and where we're thinking about doing it is essentially in a sort of error recovery or error diagnosis type state. And this all gets to like what's the safest way to deploy these things.

55:26So if I have a system that works and I have models A, B, C and D, and then someone comes to me with model Z and says it's the best thing since sliced bread, you should use it. if I put it first, well, I'm totally dependent on it working. But what I could do is actually just put it last in the sequence after all my other models. And it's kind of a throw up your hands type situation where it's like models A through D don't work, then try the foundational model and maybe that will work. And we've heard of other people thinking about it the same way. And I think really all that needs to be done is shrinking the models down.

56:03That's not super straightforward because there's this trade-off. You lose, again, some of the power of the models as you shrink them. But we're already looking at deploying these models as a sort of last ditch effort. If there's some sort of error or a totally new SKU, or there's some wild robot state that we're not sure about. It rarely happens, but it's kind of the safest place. It's like you start off one out of maybe every hundred times you try using the foundational model and see see how it performs. And then as you start to get comfortable with it, then you can sort of move it earlier in the process or use it for other things.

56:41Derik Pridmore: Well, Eric, you talked about building on LLMs and VLAs and other models. And to get to that physical manipulation, what does it mean for a robot to learn semantic control in a human demonstration? Because we're talking about these models, but again, I'm circling back to that real-world data. and as you're training them, how do you bridge that gap? How do you, whether it's a sim-to-real gap or the perception to action, what are some of the techniques that you're using? Sorry, and the question is, you're asking more about semantic control and how that differs from the previous versions? Yeah, I mean, first of all, I would just point out that for many of these terms, there aren't well agreed upon definitions.

57:38So some of what I say might not be exactly how other people think about it. So when I think, but I generally try to put things into sort of layman's terms. So let's take semantic control and call that something like getting a robot to move using just words and maybe more particularly using words that aren't precise physical descriptors. So I'm not saying something like move arm 25 degrees, you know, or six inches to the left and pick up the bottle in the middle. That's actually technically semantic control because it's words, but a lot of that is are using pretty precise physical terms that you could just parse out, you know, in a very rigid way without using natural language processing and have a robot do it.

58:37You might contrast that with something like grab the box beautifully with a flourish and turn it over to see if there's any damage and then drop it in the box. And there beautifully is a totally subjective term. Turn it over would have all of these different free variables like which way do I turn it. damage is also kind of like unbounded. It could be very specific or it could be, you know, there's lots of different things that could satisfy the constraint. So all to say semantic control to me means you're using very general words that have a lot of open-ended, unconstrained variables and perhaps even refer to even subjective, but I guess essentially invisible things, you know, like, you know, beauty or something, you know.

59:43So all of that, you sort of get for free with, and that's the power of these foundational models, is that, you know, I'm just going to continue to use, even though it sounds kind of abstract, something like beauty. There are, in these foundational models, many, many images of things that are labeled. So say it has a million images of tree and then 100 ,000 of them are labeled as beautiful tree. And it has a million images of boxes and 100 ,000 of them are labeled as beautiful boxes. It will actually start to try to build up a visual representation of what that means. And then you can begin to use that even in these kind of control sequences.

1:00:32But again, the problem is it's almost like the power of that flexibility is its own pitfall. So as you begin to use semantic control, you get more flexibility, more generality, but you lose specificity. You lose predictability. And that is something that not enough people are talking about because in warehouse settings, guarantees on what the system will do and what it won't do are very important. because you have upstream and downstream systems connected to it. So if I go to FedEx and say, hey, now this robot can do anything and it will always pick, however, you will have no guarantees on exactly how long that takes, they will not use that robot because they have an upstream and a downstream system, which if it's not producing at the right pace can just stop everything.

1:01:37And so in a home, like say, maybe we'll get to humanoids and household robotics, whether that's going to be a thing, you actually have the same requirements. You want things done in a certain way. And if someone decided to wash your dishes but just didn't wash them well, that wouldn't be acceptable. And so you're still going to have this specificity requirement. it's just a little bit of a different preference where maybe you require learning, but you don't mind how long it takes. Well, I think you're on an interesting topic here in the sense that, look, I'm an old school robotics programmer from industrial robotics days of programming moves, approaches, gripper open, all the IO sort of sequencing, those sort of things.

1:02:23And then we evolved to cobots with teach and repeat type of tablet based interfaces. And now there's the premise, several folks are predicting that, you know, the verbal interaction for the operator, not the integrator, maybe it also becomes capable for the integrators, but for the actual end user operator to just instruct the robot, you know, here's some new product that just came in from Procter Gamble, integrate these into the pick, find the pick zone and start placing them kind of thing. That's all the type of instruction it gives it, but without having to have some specific grammar to how you describe, you know, how the robot's supposed to react to your instruction.

1:03:06And I think that's the question becomes is, do you think that's the eventual end game for operators on the warehouse floor to interact with the robot's performance in daily tasks? in the other use case, of course, is what happens in the home. It's like, you know, clean up the room, put the laundry in the washing machine kind of thing, or pick up the dog toys, kind of thing, all of which have different implications for what the robot's supposed to assess and then do, right? Yes. I mean, I think the short answer is there will be some amount of just verbal interaction at deployment, whether it's the integrator, the end user, but in a production, or I should say industrial use case, because, again, guarantees are so important.

1:04:00Eventually, either that language, whatever words they're using, will be, again, sort of very specific because they want a very specific result. but then the I actually think the harder so it will certainly be the case that in the home these robots will be programmed or instructed using words but that how people build a framework to do that because of the issues you guys pointed out and we've kind of been hinting at that there's you can say something very generally but then how that gets translated into your specific house and every step of an action um has tons of choices in it there's just thousands and thousands of choices if i say fold the laundry and put it you know in my drawer there are thousands of decisions that would have to be made and and those decisions all involve trade-offs around optimization you know i spend more of my time folding the clothes that matter are they going to get wrinkled are they going to, as I put them in the drawer, get messed up?

1:05:11And so I shouldn't, you know, there's all these little things that we do as people because we're sort of always attending to something and always optimizing or learning. Whereas with robots, they're not. And in fact, it may be hard to have them always do that. So I'm suspecting that on the humanoid front, in the household, first of all, almost all of what we're seeing right now in terms of demos is either explicitly teleoperated, where I would say still today, 80, 90 % of what I see is teleoperated or pre-programmed. Or it's using what's called behavioral cloning, which is you can do behavioral cloning and semantic control at the same time.

1:06:08So you can build a foundational model and then you can take data on a robot, which is teleoperated even. So you can mix all these things together. So you can have a person teleoperate a robot, record all of the telemetry, meaning the perceptual data, the robot position. Then you can take a VLA and fine tune it on that policy. But what you'll learn, and the reason I'm saying behavioral cloning is that it will learn just that and what it will remember from the foundational model what it needs to perform that action, but it will actually forget the rest. And so what you're seeing, whether it's figures package demo or them picking up dishes, I will 100 % guarantee you that every demo you see is using a different model.

1:06:56They're behavioral cloning, using teleoperation, they're fine tuning it, they get good performance, they video it, they move to the next thing. What they're betting on is they'll eventually be able to do that in one model, learn fast enough in just one robot and be doing it all the time. The compute requirements and the performance requirements to do that, we're very far away from that. So we're actually nowhere near being able to do that. And I suspect it's going to take probably at least a couple more years to be able to try to do that. And then you're still going to have this no free lunch theorem problem where as you learn one thing, you forget another thing.

1:07:36And so I think it's going to be years before we see a robot go into a home with even a small number of preset tasks in a home that it's never seen. And then it starts to perform them and it can kind of quickly optimize, say, on the time scale of an hour to perform those tasks adequately. What we're actually seeing right now is, I think, very brittle demos that are built for each environment. And then they're kind of videoed. And it does show the eventual capability, but there's a lot of little, maybe not little, big gaps that we have to jump over along the way. And those are kind of hidden from view right now.

1:08:24Well, you know, it's not unlike the way humans work today as well. You bring a plumber in to do plumbery things and you bring an electrician in to do electrical things, right? Those are different skill sets, right, that those folks have learned over years, decades to optimize. And, you know, we sort of have this expectation that robots and AI knows everything that the world's ever known about and can instantly switch, do tasks, switch in between all those things.

1:08:52Derik Pridmore: and that even gets back to your earlier point about specificity versus generality and the assumption is oh I see a robot unloading a dishwasher and therefore that robot can do everything in the kitchen when in fact it's been trained on one task and it sounds that you're a little bit more on the skeptic or wait and see side when it comes to the dawn of humanoids in these applications. But in the meantime, do you see mobile manipulation advancing in warehouse and other industrial settings? That's an interesting way of phrasing it. I mean, I think it's a smart way, mobile manipulation versus, say, humanoids.

1:09:36There are some use cases where having a manipulator on a mobile base makes sense. But in general, we tend to talk about, you know there's sort of a few key parameters that you want to optimize in any kind of automation scenario one of them is utilization rate so if I'm going to build a machine I want it to be working it's sort of common sense you know and I want it to work as much of the time as possible so that I'm getting my money's worth for an arm on a mobile base there's obviously a trade-off there when the legs are working the arms are not working when the arms are working the legs are not working whether it's wheels or legs and that um that issue doesn't really ever go away so for for a mobile manipulator to make sense you have to find essentially environments where um you're doing a little bit of both all of the time um between a bunch of diverse sort of uh places.

1:10:47And the reality is that I've seen is that almost every environment that fits that type of description has other reasons why it's too unstructured or there's no other automation. So as soon as you start seeing regularity and something that you can automate, you start to see conveyor belts and things like that. And so it's almost like every time you can find one little task that maybe makes sense for a mobile manipulator, everything else around that task makes it not make sense, which is why people have just jumped right to the humanoid form, because in theory, it can then just go in anywhere, no matter how unstructured it is.

1:11:32So they've kind of loaded all of the problem onto, hey, we're just going to have to have super performant hardware, super performant algorithms, and then we'll be able to go anywhere. So I haven't seen one yet. I think some electronics manufacturing use cases, that's where I've seen people try this. So I've seen, you know, like Toyota, for instance, had a, had an EV factory in Japan where they were doing this. But I haven't, no, no one has seen anything productionized yet. I think the closest thing we've got, gotten to a mobile nuclear product that's, productionized is there are now starting to be a small number of humanoid startups that are forgoing the ambulatory legs and they're just using wheels, but they still have kind of two arms.

1:12:20So kind of a Rosie the Robot type thing. I think I'm almost sure that people are going to keep pursuing that because it turns out you don't need legs in a lot of places. If you're in a city, because wheelchairs have to be accessible, a robot can roll. And so I think we'll probably see quite a bit more of that. And my guess is probably, again, for safety reasons, some of the first deployed humanoids in homes will likely have wheels for a base. Because safety is something we haven't touched on, but having a 150-pound machine fall on you in your house isn't really something people are signing up for.

1:13:02Yeah, we talk about that all the time. It's one of the unforeseen problems with putting these machines in homes too soon, right? Totally. And it's not even too soon. I always say that even if, I use my even if argument, even if they figure out everything. So it's something like if it can, and this is a terrible way of saying it, but if it can slice your sandwich, it can stab you. And so it's like the more capable they get, the more there are going to be sort of perverse or dangerous edge cases. So as soon as they can go downstairs, now all of a sudden you have a much bigger problem. Because if it fails on stairs, that's a huge issue.

1:13:43In fact, our COO, Gabe Zingareti, who sort of leads our go-to-market efforts, comes from a medical robotics background. And he's on the ISO committee for the warehouse regulations for humanoids. And just, you know, what to do with stairs just totally derailed them. It becomes a huge issue because if it powers down while it's walking on the stairs, you've just got a, you know, a tumbling machine. So there's a lot of these, I think, especially Silicon Valley and, you know, machine learning researchers like to think about the algorithm side. We like to see cool demos. But when it comes to actually taking that and putting it in the home and then imagining all of the other things that have to be taken care of, there's still a tremendous, tremendous amount of work to be done.

1:14:26Derik Pridmore: So we've talked about the improvements in perception and in the models for handling diverse items, whether it's poly bags or apparel. But as the technology continues to improve, where is there room for improvement? What should we be looking at? and where should that improvement be going? That's a great question. I would answer slightly differently based on, like I would split the use cases into industrial and kind of the home. On the industrial side, I think there's a few. We can always continue to make visual perception more robust and learning faster, but to be honest, that's all already. any improvements there won't essentially make new use cases possible or really show any kind of visible improvement because robots can already essentially pick anything that's graspable by its manipulator and learn how to do that on the fly.

1:15:37I do think there's probably a lot of work to be done using machine learning powered motion planning. So having your motion plan take into account everything you possibly know about the object. You know, so if you're picking up cloth, knowing the cloth swings and that it may snag. And therefore that if you pick up a cloth, you should maybe go a little higher. or if you're picking up a long object that depending on where you grab it relative to the center mass, it will have a probability of flying off. And so that sort of stuff comes for free again with foundational models, but it's by no means optimal.

1:16:23In fact, it's sort of the opposite. It's other than the behavioral cloning aspects that it picks up just from imitating what people do, it doesn't actually really use understanding of what the object is to do something different. And I think that would be a nice, that would be a very, very nice to have if it were easy to do and it's not easy to do. So the other thing I would add is, and we've done this, but you having, and this kind of bridges both industrial and the household, having perception models that work either from both vision and force or force alone which we've done both would be it's not just very useful but it's necessary so if you've ever heard of the kind of school room experiment where they they they can put rubber bands on your hands until your hands get numb and then have you try to tie your shoes so even though you're looking at your shoelaces, what you'll find is it's quite hard because you don't realize it.

1:17:27But when you've built your control policy over time, you actually are using tactile feedback. And, you know, the dirty little truth about most of the humanoid demos is they're not really getting any force feedback. So when they're manipulating something and you see it kind of like fumble and they grab it, a lot of that is the person tele-opping, looking at it visually and knowing that they've lost the grip and then adjusting. But that's not going to be enough to get you all of the tasks that they're talking about, whether it's picking up dishes that may be slippery and you can detect, oh, wait, this is wet.

1:18:03I need to go slower or slip, you know, anything like that. So those are on the just the purely the perception side, there's still work that can be done there. And, You know, you asked about algorithms, but I would just throw in there's still a lot of improvement needed on hardware more for the home than for industrial. So in industrial use cases, what we found is although we've deployed actuated end effectors and they're useful for certain kinds of applications in a high speed industrial setting, you can just use suction. And so we get a very robust end effector that doesn't wear out. There's a problem, though, where when you start moving into these finer use cases, again, you may get more degrees of freedom.

1:18:52But what they're not telling you is that hand breaks every week and they have to replace it. And that is highly non-trivial because the force, like literally you're just constantly rubbing them. And people don't appreciate what an amazing thing the human body is and how it's always regenerating and designed for these kinds of things. So we need self-healing robots now. That's the next thing. Literally. I mean, there's going to need to be something like that because, I mean, we talk about it with our industrial customers. They ask a lot of questions about, of course, performance and what's under the hood, the algorithms and perception.

1:19:32But just as important to them, if not more, is maintenance and monitoring. So we talk about fleet operations. And just knowing things like, I should have mentioned this before with auto model, it's not always as easy as it seems. Sometimes you shouldn't be learning. So as you start to fail, you shouldn't always start learning because you may have a bad suction cup. And then you're going to start labeling what would have been good picks as bad picks. So there's all of these other kinds of monitoring that you need to do continuously on just the system itself. And I think when you think about the hardware that's going to go into home, there's at least two or three of the systems.

1:20:15So I'll say the hand and the battery need huge amounts of continued innovation before they're going to be robust enough to perform over long periods of time in a home. Well, Derek, it's been great to catch up with you. What an interesting conversation. I'm sure we could talk for another hour about all these things, and we'll have to put that on hold for the next conversation. You sit in an interesting seat in the industry, and we love hearing your perspective on what you're seeing and learning as you continue to grow and evolve your business. So thanks for coming on the show today, and we wish you all the best in the future.

1:21:00Thank you for having me. It was a super interesting conversation. I'd love to come back sometime.

1:21:08Great stuff there. Really appreciate, I guess, his skepticism. Is that fair to say of foundation models and production? Because, again, you don't always know what they're going to do. You don't always know why they make mistakes or act in different ways, right? So, again, like I said off the top of the show, that's what we're trying to do with our events, right? is, all right, there's all these new emerging AI technologies. How are the leading companies in the world that are deploying robotics? How are they thinking about AI right now? How are they using it? Why are they – if they're skeptical, why are they skeptical?

1:21:42So I thought that was pretty cool. But Gene, any takeaways for you from that conversation?

1:21:46Derik Pridmore: Yeah, of course, Eric, and we spoke about where robots are evolving in the warehouse and how AI is changing that, in some ways accelerating, in some ways, causing, again, a little bit of confusion. He said that robots can't just do everything, right? There's that whole tension between generalization and specialization. And if the robot controls make it unpredictable, then it becomes less trustworthy, and it can also affect upstream or downstream processes that may or may not be automated or that certainly rely on different systems. And so warehouses and other industrial environments really rely on the fact that, yes, we want robots to become more flexible and more capable and able to make decisions on the fly, but it cannot then introduce uncertainty into the system as a whole.

1:22:38Derik Pridmore: And we're talking here in terms of material handling processes or the flow through the entire warehouse. And so I think that was a really strong point. The other thing that he noted is that fault backs are safer than rely on one model. I'm a big proponent of having a manual backup or having the ability to revert to other types of processes or not put all your ends in one basket. You want to have a variety. You want to have diversification. That way, if something doesn't work as expected, other things or people can step in still. but he is a strong proponent of specialized models that have a better safety net than something that might be trying to do too many things.

1:23:26Derik Pridmore: And so when you're developing robots for warehouse environments, think about what is your approach? Can you work with other systems? Can you, let's say the warehouse operator decides to go into a different warehouse management system, so different pieces of software, can you make that transition along with them? Let's say that they introduce an ASRS, an automated storage system. Will you still have a place with that? Right now, we are really rapidly moving away from robots as point solutions to robots that are flexible, but that still require that performance level that is expected. And that's been one of the things I think that's held back warehouse automation is that they need to be reliable to five nines.

1:24:13Derik Pridmore: And it's that last 5 % that's the hardest to get. Yeah, Mike, we talk about this a lot, right? The underlying technologies I don't know are as important to the end user as it is to the developer, right? The thing just needs to do the job that it's been hired to do, right? Yeah, and the thing that Derek opened my eyes again to was just the tension today for robot developers, many of the audience probably completely get this, between using the broad foundation models, LLMs, generative AI for certain tasks, understanding the robot's environment, VLAs, that type of stuff, going down to very specialized models for very specific tasks, and then stitching all that together in your pipeline as you're trying to have the robot complete some task.

1:24:59And it was just another great eye-opening conversation about the trade-offs. Now, there's not going to be one model to rule them all. Every model has its specialized use case. And again, we're hearing this in other areas now. As we deploy AI on the edge, model's got to become smaller, faster, more compact, able to do something really fast without having to go back to the cloud. We were just talking about security, right? In the remote interactions, the robots have to be more functional with the chips and devices that they have on board as opposed to always going back to some big brain in the cloud to ask and answer some question.

1:25:39And I think that's what many of you in the audience are struggling with. So it was great to hear that. I love the conversation about, you know, where is the future for operators interacting with robots? Are they going to be able to talk to the robot and say, hey, pick that up, put that in that box, you know, that kind of thing, the evolution of that. I know lots of you are struggling with, you know, how that evolves. And it's really fun to watch, you know, how that's moved from teach and repeat now to watch me or just takes in some verbal interactions, descriptions in implementing them. So, Osara's on the cutting edge here in the warehouse of using AI.

1:26:15So, it was a great conversation with Derek.

1:26:17Derik Pridmore: Yeah, great stuff. Derek also talked about the need for safeguards when we're talking about that human-machine interaction, right? That a human might either deliberately or accidentally give a bad instruction and the robot has to be able to know that. right now where it's like you know you have an intern and you tell them to do something and they'll follow it out literally maybe but that could cause problems so you have to have a robot that's aware enough of the process of its surroundings as well as able to interact in a meaningful way with people and i think the promise is certainly there you know we also talk about how AI agents, whether they're embodied on board or in an IPC or even in the cloud, that they all need to have those kind of fallbacks, those kind of structured safeguards.

1:27:05Derik Pridmore: And so it's a real juggling act for the developers to make sure that they have the capable models that continue to learn, but that also have those guaranteed and structured interactions. I mean, tie it back into the cybersecurity thing. How do we know that these models are secure at this point in time, right? I mean, we're just starting to roll this stuff out, right? You're going to tell me that all these are safe and secure too? I mean, it's just crazy town. But anyway, before I get too far down that rabbit hole again, Gene, cut me off, huh? I know. quickly before we sign off here, one extra reminder about RoboBusiness coming up October 20th and 21st out in Santa Clara, California.

1:27:52Again, this is the sister show of the Robotic Summit that runs every spring in Boston. If you go to the Robotic Summit, if you had a good time there, I think you're going to love RoboBusiness as well. This is our 20th anniversary show. So we got some special things planned. We'll look back at the industry over the last couple of decades, pull out some of the bigger trends, some of the most surprising things that happened over the last two decades. But I think we have the first 40 speakers up on robobusiness.com. So go check that out, see who's speaking. There's an amazing group of people who are going to be there sharing their war stories about commercial robotics development.

1:28:28We'll have over 2000 people, over 100 exhibitors, a startup alley on the floor, lots of cool stuff. If you're looking to get involved, there's still time to do that. If you want to get your tickets, early bird registration is open for the next month or so, again, head over to robobusiness.com or reach out to any one of us here at the Robot Report and we'll help get you all set up. But that's going to do it for episode 255 of the show. Thanks again for everybody tuning in. Thanks to Derek Pridmore, CEO and co-founder of Osaro for joining the show today. Gene, thanks for being here. Great to have you.

1:28:59Again, if anybody listening to this has thoughts on this new FCC restrictions and thoughts, We'd love to hear from you. So reach out and we're happy to continue to learn about this topic and pass along our learnings to the industry. Thanks to Tiger Data for sponsoring the show this month. New episodes of the Robot Report podcast drop each week. You can find us anywhere you listen to your podcast. Please subscribe, leave us a rating, give us a review. This episode was produced by Mike Oitzman. It was edited by Tyler Kubisti. Our executive producer is Amanda Buhner. I'm Steve Crow. Thanks so much, folks, for tuning in.

1:29:31Enjoy the rest of the week. We'll talk to you again soon. Thank you.