The Exchanges

Q And so where are we today in terms of autonomous trucking? I mean, you said Otto and Starsky were the first two. There have been a bunch of other, I guess, what I would call serious attempts since then. Like, how far have we gotten?

A Yeah, many serious attempts, I would say. And by serious, I mean real technology development, real capital, real, um, real, real efforts, uh, They've all kind of followed a similar blueprint, I would say. They take existing tractors that are, that, that have been manufactured by OEM, sometimes in partnership with an OEM like Volvo, uh, sometimes, um, just sort of buying from a lot and, and, and doing a quick retrofit. Um, and, uh, you know, the attempts so far, you know, there have been a few driverless runs where, and by driverless, I mean, nobody in the front seat. There's a lot of, like, sort of debate about what driverless even is. Right. Uh, for example, I think Aurora, which is one of the largest players in this space, you know, they do what they call driverless runs, but they, they sometimes have a safety observer. This is from their public, public writings. Um, somebody sort of, sort of watching the system, but they, they call it driverless, as in maybe they don't touch the wheel or engage in any way. Um, so there have been driverless runs, even going back to that 2016, uh, auto. Um, there was a company, there's a company in Bot Auto that just, I think yesterday did a, Commercial, what they call the first commercial driverless run. Um, I saw that on LinkedIn. Um, but I would say there's not a regular driverless service for trucking on highway that exists today. Um, and …

AI assessment note: “there's not a regular driverless service for trucking on highway that exists today.”

Q camera for autonomy are more around, like, it can get obscured in certain conditions and things like that, and that's where you are Want your LIDAR or whatever, which, which doesn't have the same set of issues. Is your view, like, you can go camera only as a result of the VLMs, or is it you can lean more heavily on camera, but you still want this sensor fusion approach?

A I think, especially for trucks and, you know, given what I was talking about earlier with this 8000 pound vehicle that you might be moving on the road, you want it to be as safe as it can possibly be. And so, you know, in my mind, it's not a It's not a dogmatic debate about camera or lidar. It's what's the best technology for the moment that makes it the safest. And for, from my perspective, for a vehicle, a truck, for example, you put on the road, you would want it to have camera, lidar, and radar to do a level four, level four being, you know, driverless, to do a level four truck today. Um, and the reason, and the reason I say that is because you want to be able to see the world in multiple ways. Um, it's, but There's always a sensor doing the most heavy lifting. There's not sort of this, like, co-equal, like, democracy between the three sensors. It's usually, like, some, some sort of priority is put on some algorithms, depending on what you're doing. Um, LiDAR doesn't see traffic lights, right, very well, or it doesn't see the red-green, so you would use camera primarily for that. Uh, LiDAR can see better at night. LiDAR can see through weather in some cases. Um, and so if you want to make a safe product, And trucking in particular, you want to really take advantage of all those today. A human, and when you think about the end state, you know, a human is, is, is effectively …

AI assessment note: “you would want it to have camera, lidar, and radar to do a level four”

Q Right, because part of the market is, like, the tractor and the trailer are owned by different entities, as it stands today, right? Some, some, it's not the entire market. Sometimes it's the same player, but, you know, they get separated sometimes, and sent off in different directions.

A They get separated, and sometimes for very good reasons, like, you know, uh, trailers are relatively inexpensive, you know, tractors are relatively expensive, so there, there's a lot of, like, there's a lot of, kind of, interesting thinking around, like, trying to combine this concept, um, But just from the technology side, for example, if you've made the tractor smart, like, like, like the other autonomous truck players, and you've left the trailer, uh, conventional or like a dumb trailer, quote unquote, right? Um, you can't, for example, put sensors on the back of the trailer. Like, you have a smart tractor, but you're just taking trailers from everywhere, so you can't see behind you, ok? Uh, you can't see directly behind you. If you can't see directly behind you, you cannot Handle, for example, an accident where somebody just drives right into you. And I think that's, there's maybe some clever ways to do that, but that's, that's tricky. You cannot back into a dock, right? Because that requires some amount of understanding of what's going on behind you. Uh, one, one interesting, this is just like a kind of like an inside baseball thing, but one interesting challenge we've had in the industry is that trucks today, oh, if they're, if they're pulled over, they're required by law to deploy warning triangles, and, uh, those warning triangles go behind the vehicle. If you only have…

AI assessment note: “They get separated, and sometimes for very good reasons”

Q For those who aren't already familiar, what, what are these vision language models, these VLMs that you referred to? Like, what, what is it, and what does it actually enable?

A Yeah, so, so a VLM is like a LLM corollary is when, the way I think about it, and It's taking an image as an input and providing some interpretation of what that image is. Is it happening in that image? And, uh, it's, it's not something that really I'm, I, I had to use in my past. Like this is very new to me also. And when we started Humble and our, our head of autonomy, Drew Gray, um, has been really, really in depth in it, in the, in the VLM space, but What, what I've seen coming out of it, it's like what I think a lot of people experience now, like if you go to ChatGPT and you upload an image, right? And you say, Hey, what's going on in this image? It has a pretty good understanding of what's going on in that image. Um, and, uh, sometimes you will find intelligence baked in there or some, you know, you know, I guess intelligence is like a hotly debated concept here right now, but, but bear with me, like some understanding of the image that goes beyond there's just a dog in here. It's like, there's a dog that may have come from that Door on the left. Or, you know, this construction cone is sitting on a pickup truck, not on the highway, and therefore is probably not an important construction element because it's just being carried by a truck, just to give an example of an edge case. Um, and so, so the VLM brings some of that understanding, uh, right from the jump. And that, th…

AI assessment note: “It's taking an image as an input and providing some interpretation of what that image is.”

Q a clean sheet approach and building a vehicle that doesn't even have a cab. Um, I guess in the, in the long arc of history, of course, that's gonna, how it's, it's gonna end up, right? When we don't need a driver, we shouldn't have a space in a driver for a vehicle. But apart from that, Give me the, the thinking that led you to building a new vehicle.

A Yeah. So, uh, right, at Humble, we have this, this Cabless autonomous electric Class A truck, right? And, you know, the, the, the thinking that got me here was, is a couple things. It's kind of interesting. First, I was like, okay, if you were to imagine that long arc of future, like you're saying, what does that vehicle look like? What is the simplest possible vehicle to move freight? And it's like a box with wheels. Right, basically. So it's the, it's a platform concept where, where either a container is being loaded onto that platform, or there maybe there's a, or maybe it's just a box of the box truck moving. And that would, that would be, in theory, the lowest possible cost of moving freight, right? It can't possibly get lower than that, I don't think. Maybe there's some new, new mechanism to do it, at least for on-road trucking. So that's the first, that was the first thought that got us here, and the second was like, okay, is this possible today? Can we do this with the technology and where it is today? And, you know, that's where we started exploring it with Humble, and the, you know, the answer, the answer became like, yeah, this, the technology is there. Like, basically, there's enough There's enough here where we can take that long arc of history and move it in a little bit, move it, move it forward. Um, and then along the way, because you're doing a clean sheet desi…

AI assessment note: “the thinking that got me here was, is a couple things.”

Q that you see broadly. I presume you sort of knew from day one this was going to be an electric truck, but electric truck, I mean, setting aside autonomy even, it's got its own set of open questions on, of course, things like range and weight, Of the battery, and charging infrastructure, and charge time, and so on. So how do you think about the electric component of your vehicle?

A Yeah, and I think it, I think it starts with that, um, that vision that I talked about of, like, having fully hands-off freight. Uh, if you were going to have fully hands-off freight, you would want it to be electric to charge, to handle the charging in an automated way. It's really hard to do that with diesel. Like, how are you going to get a diesel pump You know, like into, I mean, you could maybe do it with robots. They could be challenging. So, so part of the electric story for me is just like, how do you get to this fully automated freight vision? Um, so, so electric is, is, is good from that perspective. There are challenges charging, uh, forget the autonomy side, right? Electric trucks in the U S have had a challenging rollout. They've been, the electric trucks themselves have been fairly expensive in some cases, four or 500,000 dollars. A truck today, a tractor today is Somewhere between one 52 50 depending on the, uh, on, on what you're buying. The, uh, so tractors, a very expensive electric truck. There's charging infrastructure that needs to exist. Uh, electric trucks are kind of having a moment again now because of the volatility in the world and some of the challenges that are going on, but in general, it's been a tough rollout. But what we have seen in, in, in countries like China is that electric trucks are really taking off because the infrastructure developed, …

AI assessment note: “part of the electric story for me is just like, how do you get to this fully automated freight vision?”

Q can only go You know, within these boundaries, and they do that for a while, and then they expand the boundaries, and then they go to a new city, and so on. So now we have a sense of, like, here's what's coming in terms of passenger autonomy. I imagine it can't quite work the same in trucking. So what's that, what's that gonna look like? Paint me that picture.

A Yeah, so, so I'll, I'll put Humble aside for a second and just look at the industry as a whole. And, you know, you have, you have a number of very large players, um, Very well capitalized, some public companies, all kind of going after the on-road, long-haul autonomous trucking, and they will, they will continue to work on that, and they will, there will be, you will see autonomous trucks on the highway, um, and I think you'll see them very soon, and they will be driverless. Um, the, all the, all the sort of pieces have started to come into place, like hardware and supply chain and regulatory efforts, uh, you know, that took a while to sort of make sure state and federal, um, Uh, regulators understood kind of the technology and, and how to deploy it. Um, so those large players, I think they will start deploying, um, I mean, traditionally they've been calling it a hub to hub model, right? Where it's like you sort of, you can imagine at either end of a highway segment, you know, a destination for autonomous truck to go and, and drop off rates. Sometimes you'll see it go right to customers. I don't think you'll see generalized solutions where there's just like autonomous trucks zigzagging like everywhere for a while, but Um, I think what you'll see is some specific segments, mostly in Texas, where you'll, where you'll see occasional autonomous trucks and some scale applied there. …

AI assessment note: “traditionally they've been calling it a hub to hub model”

Q weren't leveraging, like, the latest and greatest from Transformer World at that time, but obviously have been incorporating it since. And I'm curious what, like, so if you're, the difference is between, um, evolving a tech stack for autonomy that was built in 2016 or 2018 or 20 20, and then layering on whatever you can do today versus starting fresh today, like, what is the difference between those two?

A Yeah, great question. So the, the, the landscape obviously has changed a lot. And again, if I go back to that 2016 era, you know, this was even before ML had really taken off, um, in mass in at least autonomous vehicles. So, you know, to stay between the lane lines, we might just actually be hand coding. Look at these two different pixels. Is one yellow and is one not. That's where the lane line is. Like very, very simple, you know, uh, Uh, handcrafted kind of algorithms. So, so the whole space has evolved tremendously. And I think a lot of the companies have had to break down their stacks and, and re-engineer them, um, multiple times. And, uh, I'm sure Waymo, I'm sure Waymo has had to do the same. But today, uh, I'll give you one example for, for Humble, my company. Um, you know, we, most of my, most of my time in the technology space, we've been pretty LiDAR heavy, like LiDAR first, I would say. And, uh, even, even for trucking on highway, and there's, there's been attempts to develop these, like, very long range lidars that could see, you know, 304 hundred meters, so you can handle the stopping distance of a truck. The lidar technology for a long time was really good, and you can, it was a little challenging to make it, like, robust and reliable and at scale, but, but the technology itself was very good, and cameras kind of lagged behind. Like, we weren't doing so much with …

AI assessment note: “a lot of the companies have had to break down their stacks and re-engineer”

Q I have a bunch of questions for you. Let me start with this one. Um, you know, Waymo has done two hundred million miles now on public roads and is driving me all over the complicated and messy streets of San Francisco. So, you know, passenger vehicle autonomy is, is clearly here and now across a bunch of cities. And yet, we're not there on trucking autonomy. Why is that?

A Well, that's a great, that's a great first question. Um, you know, so I think we need to go back a little bit to kind of when autonomous trucking started, uh, and, and, and the history there, and it'll help explain a little bit of that. Um, so, you know, there's been like unmanned military defense experiments for a long time, but from a, from a, uh, uh, an industry that's not in defense, um, autonomous trucking really started in 2016. There was two companies Starscale Robotics and, and auto. And I, I was part of auto. I joined, I joined pretty early. And I joined auto at the time. I think there's some fallacies. I think maybe that we had at the time or misunderstandings about the space at the time, um, that'll help explain the answer to your question. But when I joined, I, I had been working on passenger car autonomy with, with other efforts. I was at Apple for a bit. And the, the challenge that I saw was Cities felt very hard. Like, uh, at the time, given where the tech was, uh, we were just like, how are we going to solve San Francisco and everything that happens in San Francisco? Um, and, you know, the way the tech worked at the time, and still for a lot of companies works this way, is we were doing this very complicated HD mapping procedure where, like, basically vehicles would drive and try to capture the world in three D, uh, and, you know, try to maintain this sort of up…

AI assessment note: “I think there's some fallacies... that'll help explain the answer to your question.”