Collaborative Robotics CEO Brad Porter on why humanoids fail commercially and what mobile robots actually do at Maersk today

Thank you for joining. What's new? Oh, fantastic outfit. Welcome to the show. Amazing. Looking great. Uh what's the occasion? First, introduce yourself, please. But uh explain uh why the fantastic outfit today. Are you doing real work? That's right. Yeah. Yeah. Hi, I'm Brad Porter.

I'm founder and CEO of of Collabor Robotics. The uh Yeah. This is how you tell whether uh whether a robotic company's really in production or in the field if they're wearing their safety vests. And uh I was at our uh our deployment with Maris yesterday and so uh had had this handy and thought uh there we go.

I'd bring it out for you guys. Yeah. So I mean to the degree that you can talk about it, uh what exactly are you doing for Maris because that sounds like important work. Yeah, we're uh we're helping them in in transload operations in moving in unloading ocean uh freight containers and loading out onto tractor trailers.

Um the you know they load carts, industrial carts uh full of kind of up to 1500 lb worth of generally boxes of product uh for retailers. Um, and then we help with the moving the carts because moving those heavy carts around all day long is is pretty pretty taxing work. And we've got You said 1500 pounds.

How big is one of those boxes? Uh, like I I'm familiar with like a 55gallon drum. I'm familiar with like a pallet of goods that you might see in an Amazon warehouse. How big are we talking? Yeah. So, think of this as as the types of boxes and and that would that would flow to a to a retailer, right?

To a big box retailer. Um and so they're unloading those from ocean containers onto carts that are about um 3 and 1/2 ft wide by about 6 and 1/2 ft long. Um and so they just load up as many as they can on the cart and then take it to you know usually these are getting dispatched out to big box stores.

Um and so you know there might be six tractor trailers that are getting loaded up to go to six different stores in a in a region. Um, so we're working um in the in the Sar Washington area, so out of Seattle port um and helping basically get distribution out to Pacific Northwest.

Can you talk about some of the differences about unloading at a port versus what uh Amazon's KA systems does within the warehouse and and some of the different challenges that you face versus what KA is doing? I imagine that some of the there's some learnings that cross over, right? Yeah.

So the way you can think about logistics is there's there's inbound flow, you know, products coming from manufacturers around the world, a lot of it coming from China. Um, and then that ends up in some distribution warehouse ready for people to to buy. So it might end up in your in your local big box retailer.

You can just go and, you know, buy a fan off the shelf. Um, or it ends up in an Amazon warehouse, uh, an Amazon fulfillment center.

So the inbound side of that is to unload the ocean containers and then bring it to, you know, some someplace where it's being stored or bring it, you know, to a to a retail store or to an Amazon fulfillment center.

And then an Amazon fulfillment center, yeah, that KA network or now Amazon robotics um what they call kind of the Hercules drives uh is their storage array. So, Amazon will have multiple mezzanine decks of um of those ka pods full of all kinds of things that you might buy from Amazon.

Literally can have a million different SKUs in a in a building. Um and then when you when you order it almost immediately, the the system knows where it has lots of those and they're stowed across Amazon's network. It quickly calculates where is the most optimal place to deliver this to you.

And then a robot goes and gets that shelf and brings it to uh a picker, brings it to someone who pulls it out of those shelves, puts it into a tote, and then those totes get routed to a pack station, gets packed, thrown uh and then it gets sorted to a truck and then usually, you know, either to FedEx or UPS or Amazon's delivery network or to USPS.

Amazon can kind of deliver into any of those outbound delivery networks. And so, so yeah, three phases coming in from the manufacturers, stored and ready to be bought and then shipped to you. How how uh where are we at in the kind of AI journey of these collaborative robots?

Um I imagine that there's tons of work that can be done with just hardcoded business logic. drive two feet forward, take a left, and it kind of just is almost like a conveyor belt on on independent wheels.

Um, versus, you know, the far future where the robot has a brain and is just making completely independent decisions and decides where it goes and pro and problem solves and reasons and and we're on the cusp of that, but I imagine that there's a there's a journey that we're going through.

Um, and so walk me through where we are in terms of that journey. Yeah. So, we we've made a lot of progress from the days where you just like followed a tape line on the floor. Yep. Right. Um now robots can can generally sense and perceive and navigate um commercial environments autonomously quite well.

usually, you know, at human walking speeds, uh, maybe a little faster, but, um, the kind of self-driving problem is reasonably well solved in, uh, in commercial spaces at those type of speeds. And that's generally done with a LAR um, and maybe, you know, stereo depth cameras.

And so, you know, LAR based slam or um is how it localizes and then navigation and planning and that can be done in a way where it can detect humans, obstacles, navigate around things.

Um, and that's the the the capability our robot has and it can do that in hospitals in um, you know, ultimately airports, stadiums, in and around people quite safely. That technology works.

that what's coming now is you can talk to robots and the robots will make the the highle plan and instruct that um where we need to get to is what we can do with our hands, right? Where you know like open up your AirPod case and pull out is a very complicated set of motions that we do without thinking about it.

We don't quite have that capability yet. Got it. Talk to me about the LAR supply chain and cost structure in your business. Um uh it's been a controversial debate point for a long time in autonomous vehicles. Um but if the economic model works, it feels like it's just pure value ad. Is the cost of LAR uh getting lower?

Are you thinking about solid state LAR coming down the pipe or is it already available? Um, are you banking on a reduction in LAR costs over the long term or does your business model just by nature of how much value you're adding, you're fine paying 50k or something like that if that's the number? I don't know.

Yeah, liars are in the kind of they used to be 50k, they're in the kind of $3 to $6,000 range right now. Yeah. Um, and so you're right, you do have to add enough value. You're not going to put that on your Roomba, right? Um, but you can put that on an industrial robot um and uh and get a payback.

Obviously, we want to keep seeing those costs continue to come down, but uh but it's not a it's not a prohibitive um element for a lot of the work that needs to get done out there.

Can you can you talk about uh form factors broadly and and what guided you your thinking towards the proxy the initial product and other products in the suite feels like you guys distinctly chose not to do humanoids even though I'm sure various VCs thought hey why don't you have you guys thought about doing humanoids demo have you seen this viral video from Boston Dynamics yeah um so I'm curious uh you know kind of the the decision-m that went into uh the current and and uh pending form factors.

Yeah. So I mean as much as like the humanoid hype seems to have been peaked in the last you know 12 18 months uh I think Elon did a lot to kind of fan that humanoid robots have been around for quite a while. Agility systems has been uh electronic these guys have been at it for a while.

And so um so when I was leading robotics for Amazon um the we we studied deeply uh humanoids in in 2018. I looked I I went through my like hype phase on humanoids in 2018. I got really excited about them. Uh what agility was showing at the time and legged mobility looked like it could actually work and it it does.

Um and so I had my team at Amazon do a full analysis of everything we weren't going to automate in another way where a humanoid could help. And I remember reviewing this paper, there were 40 diff 40 different use cases where a humanoid could be great, right?

Um and and then I looked then we looked at all 40 use cases and we said actually to solve these problems, we don't need a humanoid. Um and in fact a humanoid is kind of too complicated. You really want wheels. You kind of want to move more than three and a half miles an hour.

And is the number is the number of motors a potential issue too from a degradation standpoint? It's like, you know, we've talked to to other robotics founders that say, you know, I'm using, you know, robotic arms in my facility and we already have to replace those motors all the time.

And then a humanoid might have an order of magnitude more and actually be less productive than some type of robotic arm. So the number of motors really does drive the overall cost. It also drives the complexity of the controls, right? And so you get into a world where you need AIdriven controls.

But the problem, the real problem that people don't talk about very much with humanoids is uh getting strength out of rotational motion is very hard, right? You um you're effectively because you're just doing a short throw.

You don't have the kind of momentum flywheel effect as you get the torqus rolling on your electric vehicle, right? You're moving through very short distances. So all the power is basically your electric magnet and then your rare earth magnets.

and and so you end up needing bigger and bigger motors to get that kind of power. Um and so and then you want to wind them with the absolute highest density you can and so they end up and then you want to run them at almost the peak current that you can to get the most strength.

So, the problem is humanoid robots either have to put this way big motor that doesn't look right in the shoulder.

Um, but what they're typically doing is they're they're hand winding the motors, they're pushing the current to the max and even then they're getting maybe 60% the strength of humans and they they burn those motors out.

So, the motors are very expensive and they burn them out very quickly and they're still not as strong as a human. And so, it just it we need some breakthrough. You know, the pneumatic, like, you know, Boston Dynamics had that Atlas robot that like could do back flips and everything.

Pneumatic is 10x the uh the power of an electric, right? And but you can't you can't really make that system reliable in production. So, do you think there you think there are more consumer use cases for the humanoid form factor, maybe around the home? How do you think about applications outside of industrials?

You know, I I I have struggled to find someone someone mentioned one to me the other day that that seemed great, which is like is walking your dog. I think humanoids walking the dog would be would be quite interesting, quite cool. Um I suppose you could have a quadriped but uh um otherwise I am I am not bullish.

I do think there there's some some cool robots recently that are more kind of friendly. looked like they were kind of playing a game with a kid. Like I I think kind of the emotional companion idea is quite interesting. Um but yeah, getting the strength is tough.

I mean, even just thinking about the human arm, like the force that's generated from the human arm is from like the bicep muscle, which is much bigger than the actual joint. And so if you put a motor on that joint, you you're not humanoid with just absolute cannons.

There is there is a a humanoid company that's trying to create the muscle fibers and pull that and that's sounds like some of the pneumatic projects and maybe it'll be a hybrid. Um in terms of training and AI development there's been this talk about the sim tore gap.

I don't know how closely you've been tracking this but obviously generating data for robotics has been very difficult but now there's this new paper that semi analysis was talking about yesterday um all about training in simulation uh basically Unreal Engine you build the robot virtually you have it walk around learn as much as it can in in thousands of years of artificial data then there's going to be a gap between what it experiences in simulation and reality and and so what you do is you take what it's learned in simulation and you and you run that on a robot in a cage basically wired up with a with a power cable so that it can run forever and and it and it tries to do the moves that it learned in simulation.

It messes it up but then that generates more data that feeds back in.

Does that seem like, you know, power generation, all the all the mechanical issues aside, does that seem like an interesting path to go down for actually solving the the algorithmic and like the AI piece of understanding these uh how these robots will will actually choose what what motors to move at what times.

And do you have any experience generating data? Just kidding. I'm just kidding. Uh former CTO of Scale AI, if anyone's listening that's that's not familiar. Um, no that's so, so the challenge in robotics first is how do you get I mean it's the same in large language. What's the pre-training phase, right?

How do we get some base level in pre-training large language models? It's to kind of understand how words are likely to follow each other, right? Just statistically. Um, so motor actions, what's what's likely to, you know, to cause the arm to move forward and and things like that.

But the hard part in um in any AI system are the edge cases at the end when you're when you're interfacing with with the real world, right? And you know, fortunately, we have all this large language model data.

We have all this data from the internet to give us reference examples of what the real world of language looks like, right?

Um and so we refine on that and then we use human preference to refine even further and that's how we get you know chat GBT in in the uh in the robotics world data from simulation data from multiple robots um data from tea operation all of these are kind of techniques people are using to feed some data into you know what's kind of the pre-trained base model that gets some statistical sign correlation but when it comes comes to learning the edge cases, right?

When it comes to, hey, that doorork knob is higher than this other doorork knob or the doororknob does, you know, turns upward instead of downward. Um, you uh you and I actually selfplay to figure that out.

We come up to a doorork knob that doesn't It's funny, we have a door at Kobat that you you can either push the handlebar or there's a handle.

Well, everyone tries to push the handle and then the door doesn't open and you have to push the um and so so humans get confused too and we do this kind of refined selfplay and I think right now we're very much focused on the pre-training phase just how do we get enough data to have something that like roughly moves its hand toward the door.

Sure. To really solve this problem though, we've got to learn how to self-play in the real world um like you or I do because there's all kinds of novel stuff we're going to run into um solving real problems. Well, good luck with that.

It sounds like an easy task, but uh I'm sure you're up to the task and uh it's been fantastic talking to you. Yeah, this has been super insightful.

Yeah, I mean it's such an exciting industry because it's really just like like we're still just on the early the early part of the S-curve and there's going to be fantastic advancements. So, good luck. The future's going to be amazing. Awesome. Thank you guys. Appreciate you coming on. We'll talk to you soon.

Thanks, Brad. Thanks so much. Talk to you soon. Next up, we have Keon from Nucleus coming on uh with a big announcement. Something like 10 years in the making, close to it, maybe seven years. Uh we'll bring Keon in. Let's play soundboard.