Transcript: From TuSimple to Bot Auto: Xiaodi Hou’s New Plan for Profitable Autonomous Trucking

Executive Summary

In this episode of The Road to Autonomy, Xiaodi Hou, founder and CEO of Bot Auto, joined Grayson Brulte to discuss the invaluable lessons learned from his time at TuSimple. Dr. Hou shares his new, contrarian vision for the autonomous trucking industry with Bot Auto, a company built not on the promise of technology alone, but on a foundational goal: achieving profitability by relentlessly driving down operating costs.

Dr. Hou details Bot Auto’s unique operator model, its lean organizational structure, and how embracing the new era of AI is allowing his small team to achieve milestones at unprecedented speed.

Key The Road to Autonomy Episode Questions Answered

Key The Road to Autonomy Topics & Timestamps

[00:00] What Happened at TuSimple? Xiaodi Hou’s Perspective

Xiaodi Hou describes how after its IPO, TuSimple operated like two separate companies: one he led focused on technology, and another that burned through over $400 million trying to boost revenue for Wall Street. His attempt to shut down the “revenue company” to focus on creating a profitable product led to a clash with the board, which was worsened by a co-founder’s unrelated business dealings, ultimately leading to Hou’s termination.

[03:35] What If TuSimple Never Went Public? 

Xiaodi believes things would have been “much better” if TuSimple had stayed private, as it would have avoided the pressure to pursue an unsustainable business model that was losing $18 for every $20 earned. He refers to this pressure as a “poison pill” that forced the company to sacrifice 90% of its operating costs to boost investor confidence.

[06:54] The Biggest Lessons Learned from Past Strategic Failures 

Xiaodi’s most significant lesson is that as a founder, he is ultimately responsible for the company’s strategy. He acknowledges his past naivete in believing that technology alone could solve all problems and now sees his primary role as CEO of Bot Auto is to protect the company, with technology development being a secondary focus.

[09:50] Bot Auto’s Strategy: How to Make a Profit, Not Just Revenue

Xiaodi explains that Bot Auto is focused on creating a profitable business, which many competitors overlook. He defines true profit by considering three factors: driver cost, operating cost, and customer acquisition cost. Bot Auto’s core strategy is to maximize the “product value” that results from managing all three of these costs effectively.

[10:59] The Contrarian Approach: Lowering Operating Costs Without Premature Scaling 

Bot Auto rejects the common idea that a larger fleet is needed to lower operating costs. Instead, Hou treats operating costs as a “technology problem, not as a scale problem”. The company plans to use advanced autonomous technology for automation, preemptive system monitoring, and algorithm improvements to drive down costs without needing a massive fleet.

[15:30] Why Bot Auto is an Operator, Not a SaaS Company

Bot Auto is structured as an “operating company” that runs its own fleet of trucks. Hou argues that this is the “only model” that can achieve profitability because a company must take responsibility for lowering its own operating costs rather than offloading that challenge to a customer. This approach directly provides the “capacity” the trucking market demands.

[21:50] The Evolving Role and Cost of Sensors (LiDAR vs. Camera)

Xiaodi now prioritizes sensor reliability over having the most advanced capabilities. He argues that sensor costs are negligible on a per-mile basis (e.g., a $10,000 sensor is about one cent per mile), making the LiDAR vs. camera-only debate a distraction. Because the cost is so low, Bot Auto will continue using LiDAR as long as it helps the system.

[26:45] Bot Auto’s Testing Methodology: From Simulation to On-Road 

The company uses a statistical approach, starting with extensive simulation before moving to on-road “shadow testing” where the AI runs without controlling the truck. By analyzing “internal discrepancies” in the algorithm’s performance, they build a statistical model to predict the rate of failure and determine when the system is safe enough for a driver-out run.

[29:45] The True Goal: Cost-Per-Mile Reduction, Not Just “Driver-Out”

Xiaodi emphasizes that “driver out” is not the end goal but rather a “not so important midway” milestone. The ultimate objective for Bot Auto is to lower its operational cost-per-mile to be cheaper than a human-driven truck, as this is the only way to build a profitable, scalable business.

[33:15] A Sobering Take on Tesla’s FSD

Based on his personal experience testing FSD in Houston, Hou calls it “pretty bad,” with frequent disengagements. He suggests that while FSD might be “overfitting” to perform well in specific regions like the Bay Area, today’s AI technology does not allow a Level 2 system to achieve the Level 4 safety required for its global business model.

[38:25] Why Bot Auto is Headquartered in Houston

Bot Auto is headquartered in Houston to stay close to its operations, reinforcing its identity as an “operation company” rather than a tech company. Houston is also a strategic logistics hub with ports and a central position on the I-10 corridor, aligning with the company’s ambitions to expand east and west.

[40:20] Building a Small, Lean, and Flat Organization in the New Era of AI

Xiaodi is building a small, lean company and intends to hire fewer than 20 people after the next funding round. He believes that in the current era of AI, a small team can be highly effective by leveraging open-source tools and avoiding the need to “reinvent the wheels” of infrastructure. The company operates with a flat structure and no org chart to keep every team member focused on the commercial mission.

[47:50] From Concept to Hub-to-Hub Demo at Unprecedented Speed

In just over a year, Bot Auto has progressed to near a hub-to-hub demo, a pace about three times faster than previous efforts. This speed is credited to fully embracing the new era of AI, which allowed a small team to build the necessary infrastructure in just four months and begin on-highway testing within weeks of receiving their first truck.

Full Episode Transcript

Grayson Brulte: Xiaodi, what happened at TuSimple? 

Xiaodi Hou: Well, that will be a long story. So I try to keep it short. I think back in TuSimple days, I enjoyed myself in my comfort zone as a tech guy, as you know, like every tech guy, I feel that, , I can do technologies and technologies can solve every problem, but that is clearly, that was clearly not the case back in I feel that the company, went down to the wrong path after the IPO. I feel that, in fact, there were actually two companies inside one stock ticker TSP. One, one is a company that is managed by me doing the technology development. The other company is really not relevant to me. They’re boosting revenues run by a group of people flown in around the time of IPO. To me it was feeling like excessive personnel, excessive cash burn. We actually spend, , I think more than $400 million in terms of operating this massive fleet, trying to boost the revenue so that, you know, that was the Wall Street wants, but that might not be the right thing in terms of, the direction of the company. If I see it in a retrospective way . if you’re running a revenue company, you’re actually duplicating the operation, but for autonomous driving, what now I realize is to run a profit company, that is to build the product. And I think that’s the difference where I failed to realize I was so naive thinking about technology can solve every problem. So I was kind of like putting myself aside as a CTO’s position until it was too late. So that is really, I think, the number one part of the problem. And then, the second part of the problem actually coming from the clashes was the board. I think that my CEO term failed because the iron reconciled clash with the board. I mean, there were a recently there a, a article about the founder mentality versus manager mentality. I think that was a very good. Analogy to what I was experiencing back in TuSimple days on one hand as a founder. I have unlimited responsibility of the company’s future. But on the other hand, you can see that the board members, they have limited responsibility. That is called the fiduciary share, the fiduciary duty to the shareholders . this is where I wanted to make a change to basically shut down the revenue company, as I mentioned before, and really focusing on the profit company, just Minimizing our operating footprint, focusing on the technology development until we can make money out of this autonomous driving fleet. But, I think that was not understood by the board. However, the things were exacerbated also because, Mo Chen and Hydron, my co founder has initiated a trucking company. And, , there were some unreported monetary transaction back in 2021 before I become the CEO. And during my term as a CEO, there were some false advertisement of partnering with TuSimple, which never happened in 2022. So all of this, I would say small issues are exacerbating the distrust between the board and me. And, in the end, it was just really a, a quite turbulent time and I was terminated without cause in the end.

Grayson Brulte: Do you feel the board was not qualified to hold the position that they held? Which ultimately led to the situation that we all read about? 

Xiaodi Hou: As a matter of fact, half of the board, like there are four independent board members and the two of them, they had no board experience before. And the other two, I think each of them, they held at least six board member, , on different companies.

Grayson Brulte: There’s a lot of distraction there. I’m not ISS institutional shareholders. They obviously understand boards, and I’m not going to give a comment on on board experience. What I am curious about, though, so you mentioned two companies inside of one company, a tech company, which you clearly stated that you were leading, and then I’m just gonna call the second one business company that was focused on the business. Looking back in retrospect, what if TuSimple never went public. And you were to focus on continuing to run a tech company that eventually took over both sides of the company, made it one giant tech company without having to go through the pressure of a public market, where do you think we would be today? 

Xiaodi Hou: I think things will be much better. To be honest, because we don’t really have the pressure. We don’t really need to take the. You know, take the poison pill of. According to the Wall Street, while we are spending $20 dollars per mile to haul a freight and making $2 dollars per mile in our revenue.

Grayson Brulte: You’re losing $18 dollars a mile. It’s not sustainable.

Xiaodi Hou: Yes, it’s not sustainable. However, on the other hand, the other side of the argument would be, you know, by doing that, we can boost the confidence for the general public, especially if money is nothing. Money is an easily obtainable object back in the 2021. So basically, the company being structured by the business side as an entity that can boost confidence from the others and raising money as at a necessary sacrifice of losing 80 percent or actually 90 percent of the operating cost.

Grayson Brulte: Did you feel that the company was ready to IPO when that decision was ultimately made and who ended up making the IPO decision? 

Xiaodi Hou: Well, I mean, if you’re asking me now, I think the company was not ready for the IPO, but I think back in 2021, I was not matured enough to make any determination whether it is or it is not. And remember that at that time, I was still thinking so naively that technology can solve every problem. And I’m pretty proud that still as of today, if I’m looking back, I felt that, I’m leading the team and we have made the most amazing technology in the world back then.

Grayson Brulte: From what you’re describing you’ve grown up.

Xiaodi Hou: I hope so.

Grayson Brulte: If you’re giving an answer like that, you either have been very well coached or you’re growing up and i’m going to give you the benefit the doubt that You’re growing up, if you did not go through i’ll just use the term situation at TuSimple. Do you feel that you would be here where you are today with bot in terms of the business and technology if you didn’t have the scars from what you went through at TuSimple.

Xiaodi Hou: oh, I think you know what? I feel that, every year to me is like a thousand day back in my TuSimple days, which means that I’m growing like three times faster. Or on the other hand, I’m getting over three and three times faster than I was in the past two years, which, you know, good and bad, you know, bad is my life quality is terrible in the past two years, but the good is that I felt that I’ve grown tremendously in a lot of aspects, including corporate governance, including legal strategy, PR and almost everything that needed to keep a company, , to survive in this very dynamic environment.

Grayson Brulte: Looking back, what were some of the wrong decisions that you made that you wish that you could correct today? 

Xiaodi Hou: I think the fundamental thing is that, I am the founder of the company so that I have to take care of the strategies of the company. I think overall, uh, the biggest lesson that I have learned is not about anyone else. It’s really about myself. I feel that now I can say that strategy is a big amplifier. It can either bring huge value to a company or huge damage to a company. But, you know, if you look at today’s world, there’s a lot of strategy failures happening in the world right now. And, very interestingly, it’s nobody’s fault. Like if you look at today’s Intel or Boeing or many other companies. Big and small company that you know, I’m naming this big companies only because they’re famous and that they have succeeded tremendously. So that’s why their strategic failure can be actually known to the world. But big and small company, we all make strategical mistakes. But this strategical failure, even viewed as nobody’s fault, they are in fact, the founders fault. Especially for startups. So I think this is the biggest lesson that I’ve learned. And I, I have no excuses for the strategic failures of As I said many times, I’m, I was naive about technology and I think technology is not going to save the world alone. They need to be protected. , so this time I become the CEO of the company and, , my primary job is to protect BotAuto. And my secondary job as a hobby is to build more technology, contributing on the technology side.

Grayson Brulte: When you look at Intel, you brought Intel very famously Intel missed mobile. They completely missed the boat on mobile. And we saw what happened with Qualcomm there on market share. Autonomous driving is changing. You have your PhD from Caltech in computer vision as CEO, a bot. How do you avoid making that Intel mistake when the new technology for autonomous driving evolves and is ready and bots ready there to capitalize on it and not look back and be Intel? 

Xiaodi Hou: Well, I think there are a lot of things that needs to be answered, for Intel’s strategical failure on missing the mobile side, as you mentioned, I think this is more about the reading the trend of the future. And I feel that, the uncertainty in terms of the strategy for Bot Auto in the future is not about the future trend. And by the way, simply because that we have seen a very clear trend in the future, which is an undeniable shortage of the truck drivers, which is exacerbating every day. And that is the reason why I am so dedicated to join the autonomous trucking industry again, But I think the problem, , or the strategical uncertainties for BotAuto in the future is that how can we structurize a roadmap so that we are focusing on the right thing? That is how to make money.

Grayson Brulte: How do you build that roadmap? 

Xiaodi Hou: Okay, maybe let me show you some failed examples. I mean, some other companies in the autonomous driving field, everyone is talking about autonomous driving as a technology feat. People thinking about is like, wow, you look this truck and run by itself. How cool is that? But, as a company, many of us has failed, including TuSimple’s past has failed to answer the question of how to make money, make money means making profit, not making revenue. Right. And in order to gain profit, there are actually three parts. One, the driver’s cost. And the two, the operating cost. And three customer acquisition cost. People are thinking about drivers cost people are basically advertising about the driver cost reduction, but they’re omitting about operating cost and customer acquisition cost. But in this round for bot auto, I need to strategize amongst all three because all three adding up together driver cost minus operating cost minus customers acquisition cost is the product value and how to maximize the product value and how to minimize the time while we’re developing the product value until the product value is positive is really the key to success this time.

Grayson Brulte: How are you going to lower those costs when you eventually enter commercialization? 

Xiaodi Hou: So here’s the thing. So many companies thinking about a strategy of, you know, what operating costs would be lowered if I have a bigger fleet. That’s kind of like a non brainer, no brainer answer, right? But this is actually wrong. The right strategy where, I mean, this is probably a contrarian strategy that I’m having right now in Bot Auto is that, can we lower the operating cost without premature scaling up. And the answer is yes. So we wanted to think about the operating cost as a technology problem, not as a scale problem. That is, we are going to use the Iterative new autonomous driving technologies to lower the operating costs. For example, automating everything that we could and do a lot of preemptive monitoring on the situation of the hardware and software health and also improving our algorithm so that we have less dependencies on this infrastructure and software that are very hard to maintain. And by removing those, I think the overall operating costs can be reduced one day to the point that is lower than the driver’s cost.

Grayson Brulte: Are you following this was pioneered by Lufthansa in the airline industry preemptive maintenance? Are you following that were you using sensors to preemptively? Get of the head of the maintenance costs, because in the old days, if you have to go tear, let’s just use a Boeing 747, but tear down cost time. It takes a lot of time and cost setup. And when they went to preemptive maintenance, there’s sensors and then they would know, okay, part two and the right part of the fuselage. And they could fix it there. Is that a similar model that you’re following that airline model? 

Xiaodi Hou: You know what? I have some like I have some homework to do after this. I’m going to go read the Lufthansa story. But I can tell you. My epiphany of this preemptive, maintenance is really coming from my. My experience after the driver out. And you know what, as a matter of fact, many of the other autonomous train companies, they have not finished any driver out. And there, I think they’re really in the struggle of delivering their own driver out. But for us, while we were doing our driver out, the thing that I found is that. Okay. First of all, there is a line of absolute safety so that we need to make sure that all of the, all of the sensors, all of the hardware, a lot of software needs to be up and running in a super healthy level well before every launch of the driver out run. And in addition to that, all of the redundancies also need to be 100 percent prepared before the run. And that gives the maintenance team and nightmare. It’s just like if I’m a normal person, I’m just going to work every day. My bio status like my health level can be, you know, in an average normal healthy level. However, if you consider me as an astronaut, I’m going to Mars tomorrow. Then you’re doing a health check for me. You will find probably 25 yellow flags on me like my health level is not ready to go to Mars. So, , the level of healthy, healthy check is very different when you’re doing the normal driver in run versus a driver out run. And that is one of the biggest, biggest sink for all of the costs for driver out of autonomous autonomous runs. And this, after realizing that point, I have known I have known what the focus should be that is to operating the up lower the operating cost by automating everything that is specifically designed for a redundant system. And the today’s system is not ready for that.

Grayson Brulte: How do you get today’s system ready for that automation future? Because automation is going across every aspect of industrial that we can imagine.

Xiaodi Hou: It’s really a large number of small efforts. I just give you 1 very, very, very basic example. Like, are you sure that the sensor is clean enough before you go? very And of course, there are maybe five different ways of checking whether the sensor is clean. One of them is to let the test operator do the check before everyone. But if you put that thing into your, your cost model, you will find that they’re wasting five minutes just to check the sensors. Five minutes, it’s just money. So you wanted to avoid that, that five, five minutes of wasting. You do some automated algorithm to improve that. And, something else like, you know, you’re checking the health level of your SSD. You’re also checking the health level of the cooling system. Everything.

Grayson Brulte: Is automation, if you want to call it, one of the core pillars of bots business? 

Xiaodi Hou: I would say automation is a method. The goal is to lower the operating cost. And lowering the operating cost is not one of the pillars. It’s the pillar. The monolithic pillar of Bot Auto. I can probably tell you this. We are an operating company.

Grayson Brulte: okay. You said it. I’m asking you’re going to operate the fleet then.

Xiaodi Hou: Yes, we are.

Grayson Brulte: How’s that going to work? 

Xiaodi Hou: I think more fundamentally, I would ask everyone else if you are not operating the fleet. How’s that going to work? There is no way for you to, , in the most efficient way, lowering the operating costs. If you’re asking your customer to lower the operating cost for increasing the product value for a product that you provide, that’s not fair. You see that the problem really here is that you have to lower the operating cost by you because that’s a product that you offer to your customers. And of course, by saying this, like hundreds of investors will challenge me. Oh, then shall the your in , , asset heavy business. That’s poison. My response is like this. First of all, we are an operation. Heavy company. We want to be asset light. But asset light, it’s an option asset heavy is not the end of the world. Yeah. The key thing here is that you need to make a very strong thesis on how to make money. Everything else is a kind of a secondary thing.

Grayson Brulte: do you feel this model is the best way to achieve margins that you’re trying to achieve? And if so, what margins are you trying to achieve? 

Xiaodi Hou: I believe so. I believe this model is not only the best model, but the only model that can achieve the net positive for the product value.

Grayson Brulte: Do you have any estimate of the margins that you’re trying to achieve? 

Xiaodi Hou: I would say the net margin once we scaled up is probably on the range of maybe 30%. Including everything.

Grayson Brulte: Would you have been able to get to this model if you didn’t go through what you went through at TuSimple, especially with the ghost rider. When you had the driver out, would you have been able , to get here from a model perspective? 

Xiaodi Hou: I think the TuSimple Ghost Rider experience gave me a lot of, uh, reflections about the things that Could go wrong and, , for now, I’m actually building everything from a new and in fact, , a lot of things has changed. The ecosystem has changed, the hardware that are available off the shelf are also changed and their failure rate has changed. And also the the open source ecosystem of the machine learning infrastructure has also changed. So given all of these changes, I think it’s probably more capital efficient for us to forget about every details, but only remembering the principle of operation first cost operating cost reduction being our sole mission. And then we can just rebuild everything from scratch. I think by doing that, on one hand, we don’t have any technical managerial debt. But on the other hand, it’s really, you know, a fun experience to embrace that whole new world in building things.

Grayson Brulte: Do you feel the market is ready for the bot business model because your model is different from every single one of your peers in autonomous trucking? 

Xiaodi Hou: Yeah, I realized the difference. And, you know, that really gives the heart, gives me a hard time when talking to investors, especially in the Bay Area where, you know, everyone is talking about SaaS. So my model is being criticized by a lot of people, but, you know what, , if this is, , if this is the truth that I believed, I think I need to defend it. And I don’t mind of defending this, because on the other hand, , our model, even though it’s different from everyone else, it’s actually compatible with the market. I think one of the thing is that, , maybe this is a little bit digression, but I wanted to say it anyway. , the funny thing is that in Silicon Valley, I’m going to use a Lord of the ring quotes, history became legend, legend became myth. So people forget about the actual first principles. They all really follow on those derivative concepts. So SAS is a perfect concept. On one hand, if you’re talking about the market. What the market asks the market for autonomous trucking. What would they mark? Okay, so let’s don’t talk about autonomous trucking. The trucking market only asks for capacity. I mean, if you go to the trucking market, nobody asked for nobody’s craving for autonomous driving technology. That’s not the case. Everyone is asking for a hey, I have a load. I need to I need this load to be shipped tomorrow or today. Who can help me? That’s the only ask by the market. But people are taking derivative concepts of saying like, you know what, if I have a software, someone else magically or heroically is going to reduce the operating cost for me. And maybe a third party is going to connect me with some shippers. And the fourth party is going to, , making sure that everything worked together. And then boom, we, we launch our product. But I think that’s is really a deriv derivative concepts not real, not the real thing. The real thing, a very straightforward thing is that the markets needs for capacity and what bottle auto provides is the capacity. So we are a direct correspondence to the market, but I think this, this very raw, direct way of connecting to the market is not being appreciated by some of the investors. At least for now.

Grayson Brulte: You’re listening to the market. There’s a demand for the market. And looking at the market, you have the nuclear verdict issue. You have the rising insurance cost issue that’s going across the entire trucking industry, not just autonomous trucking. What impact Will your model have on insurance costs? Will you be able to have a lower insurance cost for your model than your competitors in the field? And if so, can you then you pass on those cost savings to your bottom line to increase your margin? 

Xiaodi Hou: well, first of all, in terms of the cost per mile, if we’re taking cost per mile as a kind of a golden rule for us, then the insurance cost is we’re about the same as human trucking company. Right now. And of course, in the future, if we scale up, the insurance costs will probably first go down, depending on our negotiating power . but the insurance cost is relatively small comparing to other parts of the cost. Remember, we’re talking about $20 dollar per mile as the beginning point, and we wanted to reduce that $20 dollar per mile to somewhere around $2 per mile per mile in total operating costs. And insurance costs like if we can have a good bargain, maybe we subtract maybe 5 cents out of it, but it’s really not the first priority for us.

Grayson Brulte: How muchh in the cost per mile is the sensor sweep? Because when you built, let’s call it TuSimple, we’ll call that round one. The sensors were very, very extremely expensive in some cases now, what’s called bought round two, obviously, we’ve seen the market do it. Mr Market do his job and the sensor costs have come down. What impact is sensor costs having on the per mile? 

Xiaodi Hou: Previously in my last life, I would, , would argue for the most capable sensor. , like who has the longest range, what has the most, , who offers the most functionalities. But now what I’m looking at is the reliability of the sensors, because if you’re thinking about this, if one sensor can last a million miles and then the per mile cost of the sensor is minimum, like say a sensor, let’s just randomly say a $10,000, , a $10,000 dollar sensor is only about one cent per mile. I don’t really care about that one cent per mile when it comes to CPM optimization, CPM meaning cost per mile optimization, right? So I think the sensor cost is really minimum. And right now we also admit that the more we buy, the better price that we have. And right now we have only two trucks operating right now as of today. So we don’t really have that bargaining power, but we don’t really care too much about the sensor cost. We know it’s not going to be a major issue for the cost per mile journey.

Grayson Brulte: How do you see the center suite? Changing and do we ever get to camera only for class-8 autonomous trucks? 

Xiaodi Hou: Again, I think this is another, , thing. If you really look from the cost per mile perspective, you realize that once in mass production, the sensor, the LiDAR cost will be probably 0. 1 cent . per truck per mile. That is really like negligible. And then it’s really, I feel that the argument between LiDAR and the camera only or LiDAR and plus camera and camera only is really the, I would call it a technological ideologization. People wanted to criticize each other by oversimplifying the problem, but What is the fundamental metric that you’re looking at that is cost per mile? If you look at cost per mile, it’s trivial. We don’t need to argue against each other. If you think it will help, buy it. Because it’s less than one cent per mile. So my, my prediction, just to directly answer your question, I think, in my company, at least in foreseeable future, we are always going to use LiDAR. Because it’s so cheap.

Grayson Brulte: I’m thinking about this. Let’s say you and I, we take a walk and we go visit a truck dealer and there’s a traditional diesel powered truck. You look at it. I look at it. I’m thinking one thing. You’re thinking another. When you look at it, how do you think about making it autonomous? What goes through your head when You’re looking at design, functionality, range, cost. What goes into that when you’re looking at a truck and you say, I’m going to make you autonomous.

Xiaodi Hou: very good exercise. So for me, number one is that do we have a redundant breaking redundant steering or the capability to upgrade the truck to make it redundant in terms of breaking and steering? That’s one. And then the second one is that I think the internal wiring in terms of harnessing and also the redundant power, all kinds of wiring and all kinds of like double E components is, is it going to provide us a kind of a clean and safe solution so that we can make every wiring redundant. Other than that, , we may thinking about, you know, how to install a sensor rack onto this vehicle chassis. Once that is done, I think it’s probably done. I mean, we don’t really have too much of the actual requests for a truck.

Grayson Brulte: Are redundant chassis common in the marketplace or is that a bespoke product that you have to have specially made just for Bot Auto? 

Xiaodi Hou: Well, there are, in the early days, we certainly wanted to have a lot of, , a so called retrofitting process. Because nobody, As of the main truck manufacturer today, they’re building the exact model that we want for all of the redundant wiring and the additional sensor rack. So we have to retrofit early days. But one of the interesting observation that I had is that retrofitting doesn’t mean low quality. Retrofitting means it’s more expensive, but it doesn’t really necessarily imply low quality. If you I mean, if you if we find a qualified tier one supplier for the wire harnessing for the all of the sensors and actuation component like steering and braking. And it actually worked. So this is my past experience. But if we’re talking about the cost, then the operating costs. We’ll put the cost per mile hat on and we find that the additional cost of retrofitting is again negligible. So that’s why I’m still confident in early stage retrofitting, later stage when the cost can just, where the scale can justify the cost reduction. We’re going to negotiate, of course, with an OEM to, to build, to massively build a large number of trucks.

Grayson Brulte: Before we get to the OEM, you have to go through various different forms of testing. You have on road testing, track testing, simulation. How do you know when you make the decision to be to go from simulation from track to on road testing? 

Xiaodi Hou: This is the statistical problem. First of all, we, we certainly employ a lot of simulation testing. And in my opinion, simulation testing is really helpful in the early day of testing. But when you’re accumulating enough simulated, miles were simulated testing experiences on road testing is definitely needed. There are several aspects of on road testing. You can either do so called shadow testing, where you basically, let the algorithm do everything except for actuation. Or you let the vehicle just do the actuation. So most of the testings are providing information to our system. On one hand, in the shadow testing, we can see the discrepancy of the behaviors of the human driver versus a computer driver. On the other hand, of course, doing the testing with a human driver behind the wheel is going to tell you that the real performance of it end to end. These are all needed and based on the testing results we can build a statistical model on how often do we encounter a problem, you know, behind every accident, there will be maybe 10 close calls. And behind every 10 close calls, there may be 10 or even 100 discrepancies in terms of the algorithm internal representation. So we’re really harvesting this internal discrepancies as an indicator for us to predict the rate of failure or the rate of safety critical failures. And then when we have enough confidence, we feel that, yeah, we’ve solved most of the problems, it is probably going to be safe for a couple million miles or even more to be driving on the road and then we launch the vehicle.

Grayson Brulte: Is that when you, when the statistical model lines up, is that when you make the decision to then go driver out.

Xiaodi Hou: Yes.

Grayson Brulte: So, in theory, you could become the first person ever to do driver out with two different companies, two different stacks, if I’m reading this correctly.

Xiaodi Hou: I think so. Yeah, I mean, I don’t know about the other companies timeline, but it seems that everyone in this industry is having a tendency of delaying their, their promise milestone. I, on the other hand, I’m the hardcore guy. I never, I never delayed one of my past promises in my career, which I’m pretty proud of. So hopefully I can, I can also be the one that, launches this again before anyone else has launched it in public highway.

Grayson Brulte: Is driver out one of the Ultimate goals for bot. I’m not going to ask you for a timeline because I think timelines are inappropriate at this time, but is that one of the ultimate goals as you go forward and you continue to build bot.

Xiaodi Hou: Not at all. This is where I actually, you know, I pivoted from a CTO to a CEO. I think driver out is more like an internal testimony to our own technology and how things are working together. And by the way, it is very important to set a very clear goal to mobilize the team. But this is by no means the end of our development. It is actually. A not so important. Midway for our final goal, and our goal is as always this, making the cost per mile of our truck operation to be lower than that of a human. And only by achieving that goal, we can make money. We can make every truck a money printing machine. And only after that, we believe that we’re qualified for a scaled operation.

Grayson Brulte: when you achieve the profitability per truck, before you go to scalable operation, as you said, what do you have to do from a safety perspective ? 

Xiaodi Hou: Safety is an uncompromised, threshold that we always have to maintain. That is why, as, as I said, we are facing enormous amount of maintenance cost when we launch every driver out That is, in the early stage, seems very unpractical. Like, why are we doing these many checks? Like, we’re spending hours before every single run in the early days of a driver out operation. You never, you will never commercialize it. Of course, we hear it. In the early days, we will not, we will not compromise this safety standard, even though they’re making the process, like, insanely complicated. You never, you will And then the second step is that we’re going to simplify this insanely complicated process step by step with our technologies.

Grayson Brulte: and by simplifying, is that where we talked about earlier? The automation comes in as the technology evolves.

Xiaodi Hou: Yes.

Grayson Brulte: Then since you’re taking, I’ll use the term, a new approach, then do you have to get outside validation by a, an independent third party that gives you the check marks each way? Or do you just go about it your own way? 

Xiaodi Hou: Oh, definitely. We are going to talk to a lot of external parties. And in fact, we are talking to external parties, which I cannot name their name under the NDA. But yes, third party validation. We love it. And we want we are confident that our methodology is compatible with their safety principles. And, in fact, we have already had some conversations, and I just feel like we should meet earlier and we should have a beer together because our philosophies are so aligned with each other.

Grayson Brulte: Well, that’s refreshing. You’re taking a, I’ll say, I’m going to call it, you’re taking the Xiaodi Hou approach , to building autonomous trucks. It’s, it’s a different strategy this time. You’re going to start with the two trucks that you discloused. Once you get past that, you meet all the certifications, it’s called, you get your green check mark of health. How are you planning to scale the bot business while also keeping your attention focused on the bottom line, which is cost.

Xiaodi Hou: well, I think the whole thing is that we are using the minimum number of trucks at every stage. However, after our stage grows, the minimum number will increase because we will need to, have more exposure so that we can fix or optimize the operational issues. That is going to be a very natural process. We’re going to probably start with less than five trucks to 10 trucks to 30 trucks. And at some point around the 30 trucks or 100 trucks, we’re going to break even, which meaning that our operating cost of our autonomous fleet is going to be lower than that of a human fleet. That is my estimate, less than 100 trucks.

Grayson Brulte: is your advantage in the marketplace, purely your cost structure.

Xiaodi Hou: I think that’s the only validation that , we need to, , go through.

Grayson Brulte: Then you go to that validation. Then there’s a little company out of Fremont now Austin, Texas called Tesla. They just put on a Disneyland event over Warner brothers. No talk of the semi. But there is a lot of talk that FSD will come to the semi. What are your thoughts, one, on Tesla’s FSD approach, then what are your thoughts, two, on FSD going on tesla’s semi? 

Xiaodi Hou: Oh, wow. Okay. So you’re talking to the right person because I’ve been bombarded by the same, very same question so many times. When I got a chance, I actually, I, I recommended to my wife to replace her car to a Tesla Model Y. And I was actually testing the FSD every week. Of course, the newest version of it every week and my experience in Houston, I don’t know what, , is pretty bad. I mean, disengages every 10 minutes or so, and it’s a, it’s kind of like a trick that I will reserve for the investors. Whoever comes to visit us in our Houston facility. I took a Tesla. I pick up the investor from the airport. I start FSD . after the trip, , they will have no question regarding Tesla. So, but I don’t know, you know, it seems that the, Bay, the Bay Area, the Silicon Valley is, fantasy world where FSD is working much better in the Bay Area, which I’m not surprised because, for any AI, there is always a, I would call it overfitting. Principle, like if you overfit something, it will certainly get better in that specific domain. But on the other hand, if you look at the FSD’s business model or Tesla’s business model is to sell vehicles, they wanted to sell vehicles to everywhere in the world. So the product that they offer certainly needs to be, all applicable to the entire us or the entire world. And that is giving a very different level of safety requirements. For autonomous driving system or let’s just be clear. FSD is a level two system, not the level four system. But the thing is that if you’re, if you’re trying to develop a level two system that has to be applicable to everywhere in the world. Today’s technology in the world that does not allow this type of wide applicability to reach to the safety threshold that is required for a level four system.

Grayson Brulte: Do you think he gets there? 

Xiaodi Hou: certainly they’re not getting there, at least from my experience in driving in Houston with FSD. That’s on one hand. On the other hand, is that, they are operating. They are actually selling vehicles. However, if they become an operator themselves, they’re actually competing with their own customers. This is especially true for trucking.

Grayson Brulte: If you look at, I keep using the word Disneyland, because it was, to my opinion, it was inspired by Mr. Walt Disney and his vision for Disneyland, what Mr. Musk did at warner Brothers, they’re going to sell you, myself, the cyber cab and we are going to operate based on their historical precedents, that they are going to sell the semi and then sell an FSD license on top of that. If they do, this is a hypothetical now, if they do, and they sell the license, does that model then come into the more traditional autonomous trucking side? Or so, okay, we’re gonna sell you this plus an add on, does that model sort of translate over because the market might be accustomed to it? I said that market might be accustomed to it because Mr. Musk is extremely well at marketing and getting a lot a lot of attention for things.

Xiaodi Hou: You’re right, but I think the key distinction is that, , who is going to, who is going to be , the faulty party, , if there’s an accident for every level four operating company like Bot Auto, if our autonomous vehicle had an accident in autonomy, it’s on us. But I think right now Tesla’s model is that whoever has the, like a . safety incident is on the driver, and I don’t see that part changing in near future. I think that’s a distinction. Basically, we’re selling different products.

Grayson Brulte: There’s a lot of insurance questions to be asked and you’re right. You are selling different products. There’s there’s no denying that and If you look at the historical way that FSD is rolled out on the vehicles, they’ll operate quote unquote, I’ll use air quotes, anywhere. Obviously, I’m under the assumption, based on historical precedence of your competitors, that bot will operate on lanes. Is that a fair assumption that you’ll start with a lane by lane model? 

Xiaodi Hou: Yes, I mean, because I have a very clear understanding of the capability of today’s AI technology. Even though I mean, from the outsider’s perspective, AI is advancing every day like a tremendously, but fundamentally, There are some very deep constraints that, you know, remains unresolved as was as if it were like 10 years ago. So I, I think I have a very clear, conscious understanding of today’s status of AI and my attitude towards autonomous driving is that, yes, I know it’s just basically telling everyone that we’re over fitting a lane. And I’m probably overfitting a lane because of the commercial makes sense.

Grayson Brulte: So you want too be the Savile Row tailor for the lane where you have a beautiful bespoke fit.

Xiaodi Hou: Yes.

Grayson Brulte: Will Houston be The hub, or if you want to use the term depot for bot operations, where the trucks will, they’ll live in Houston, they’ll go out on their lanes and then they’ll come back at the end of their journey.

Xiaodi Hou: We can have other hubs, but Houston is definitely the headquarter and the biggest hub of bought auto operation, but we can branch out from Houston to San Antonio to Dallas, east, west, north.

Grayson Brulte: All the rest of the industry is up north in Dallas and you’re in Houston. Why Houston? 

Xiaodi Hou: First of all, you see the other company they are operating in Dallas, but none of them is headquartered in Dallas. But we, as we, as you know, we’re hardcore guys and we headquarter ourselves in Houston. There’s a reason for that because we aren’t an operation company. We keep telling ourselves we are not a technology company, even though we are confident that we possess the best technology in the world, but we do not treat ourselves as a tech company. We want it to operate. We want it to get closer to the trucks, even at an expense of getting farther away from investors. And, well, I mean, there’s a, maybe a Houston and Dallas, Discrepancy, but I would consider multiple factors. For example, Houston is a city where they also have ports, and also Houston is closer to the border where a lot of, , cargoes from Laredo, they have their sorting centers in Houston. I just love I-10 corridor because of my past life. So I think, starting at the. Exact the middle point of the item corridor. It means our ambition to go through the east and the west.

Grayson Brulte: From a talent standpoint, cause you’re going to have to hire very smart men and women to come work for bot. You have NASA there. You have the, the oil and gas industry. How’s the AI talent in Houston? , is there a large group of talent for you to pull from? 

Xiaodi Hou: Oh, by the way, I’m glad that you mentioned NASA. I mean, this is also a kind of a romantic thing for me personally. I think this NASA spirit is really, resonating in Bot Auto, we wanted to pioneer the unknown. And that’s why Houston is, you know, a beloved place for this. We share a lot of spirits here, but in terms of talents, first of all, yeah, most of the talents are has to immigrate to Houston. We don’t really find that many local AI talents here, I have to admit, but is it a big issue or is this a small issue? I would say it is a relatively small issue for two reasons. One is that wherever we have the talents. They are all over the place in the U.S. They have to come to Houston for operation purposes. Operation is the king. That’s one. And the second is that how many people do we need? In fact, another lesson that I’ve learned is that we’re now living in a new era off AI where we do not want it to kind of over saturate the talent. For this company. And I’m going to build a small company. I can tell you, you will be probably surprised after, even after this round of fundraising, we’re going to hire less than 20 people . we have about 45 or I think 48 ish, but, we’re going to hire less than 20. I think till the first half of next year, no more than that, we don’t have the headcount. The reason is that I think in the new era of AI, it is actually possible. To build a small company that is focused on the product rather than focusing on a lot of the periphery heavy lifting tasks such as big data, deep learning, training, infrastructure, and all the other stuff we can focus on the product. Imagine that there is another company Midjourney, very famous, right? They are famously for also having a very small team. And I think that. It would not be possible for Midjourney to have a small team five years ago because all of the infrastructure were not ready. But, now we don’t need to reinvent the wheels. We can, we can just acquire the wheel, the wheels from the outside world. And then even in a, in a open source domain, we can just acquire the wheels very openly. And, that allows us to focus on the, on the core thing. That is the, the product development. And also, maybe another interesting observation is that for deep learning, years ago, we used to have a multiple neural networks existed in our system, like a camera detection is one LiDAR detection is another than fusing together and then tracking prediction is one of them is a dedicated in your network. But now, you know, with the events of transformer and a lot of other new neural network architectures. I think the tendency which we certainly embraced is to have one gigantic neural network that, that solves all of the problems, either you call it a foundation model for autonomous driving. But again, I don’t want it to be considered as a mean, ideologization of technology. So I wouldn’t really make a slogan out of it. But I think what we’re doing is to use one gigantic neural network. That has a lot of shared components, shared weights, to solve most of the driving tasks, , in it. Of course, there are some other things that we also do outside the neural network.

Grayson Brulte: If you look at What’s App pre acquisition by Facebook now known as Meta, that was a very small team. He said, you’re right. Midjourney is a small team. We’re starting to see a trend of small teams do really great things because the compute can power the access to resources. And you’re going to be after you close your funding round, you hire, you’ll still be under 70 employees. That’s very small in the grand scale of things. Do you see small teams solving great problems continuing as a trend? 

Xiaodi Hou: Well, I think this is, again, we’re, we’re stepping into the uncharted territory. , I’m actually exploring a new way of organizing a company. And that is the A.I. way of solving the problem on organizing the person Very interestingly, I can show you one thing that, sometimes candidate asked me, what is your org chart and tell them? No, we don’t have an org chart. That’s quite interesting because I feel that, everyone is collaborating with one another. Using some sorts of Protocols. And previously, these protocols are maintained, with a labor intensive manner, such as the PMs are actually, , joining different team together. But right now, if everyone is sharing their experience using the same kind of foundation, either in that using one foundation model to share their They’re they’re working product, their yield, or they’re using the more generic way of looking at things by defining things in the in the less articulated way in a less siloed way, which is more of a today’s AI’s way. I just feel that the team boundary is getting very blurry right now.

Grayson Brulte: So you’re going to just be a complete flat organization with no hierarchy.

Xiaodi Hou: Yeah, I think it’s like a manageable chaos scenario for us.

Grayson Brulte: But that’s going to allow you to have Your hands on different aspects of the organization. You’re going to as you described earlier clearly implement the lessons that you’ve learned and you’re going to go to the memory and the scar tissue to build a a company that’s focused on profits. When are you planning to launch commercial operations, both with the human and then eventually when are you planning to move towards driverless operations? 

Xiaodi Hou: well, I think these two things are the same. And, maybe I’ll just go back for a second about my previous statement that having a work chart, less organization is really helping us to make every algorithm engineer to be aligned on the grand vision on commercialization. Whether we have a driver, we’re not having a driver. It should make very little difference for a deep learning engineer because all that deep learning engineer needs to focus is on how the deep learning algorithm can contribute to our grand campaign of CPM reduction. And that can only be done if the most techie side of the company is connected with the most operational side of the company. So that’s why I think we just. Break the wall. We don’t have any operational silos, or we don’t have any organizational silos. And all of that is like, either you have a human in the truck, we’re not having a human in the truck. The protocol is the same. So I think we will remove the driver whenever we feel that the statistical numbers give the green light. And of course, this statistical approach is not going to be validated only by ourselves, but also by some third party as well. But, , I think for us, we will treat , as if we were having a driver out run and, , on one hand, on the other hand, we’re also treating us as if we were hauling this cargo for commercial purposes.

Grayson Brulte: Is the algorithm team learning from the operations team and vice versa with a clear goal of, I’ll use no pun intended, driving the truck forward? 

Xiaodi Hou: Yes, and now our, , truck facility is like , 20 minutes away from our office. So it’s very convenient for everyone to ride on the truck. Sometimes you know, there’s a configuration issue that waste 20 minutes. You know, the algorithm engineers will feel guilty about this 20 minutes waste. Right. But previously, if your, your algorithm is, is like a sitting in a, in a beautiful office in the Silicon Valley where the testing is done somewhere in, Dallas, make it, and then in that situation there, I think there’s a strong, condescension, you know, the algorithm engineer is like, no, you’re a stupid, like you should offer, you should configure it in a certain particular way. So that you don’t really waste that 20 minutes. But now, if you’re really in the in the facility floor, that’s a peculiar configuration. It’s actually because of your algorithm is not designed in a compatible way of day to day operation. So people would be owning more of the operation risk or operation failure to themselves.

Grayson Brulte: In theory, that would allow you to scale more efficiently and more quickly.

Xiaodi Hou: In practice, we are, you know, we are one year and two months old and our trucks almost, uh, up and running for a hub to hub demo. This is probably three times faster than the world’s second, second quickest, , hub to hub demo, which is also created by my past company, but we’re very proud that this organization actually worked.

Grayson Brulte: Let’s just, so I’m going to say sub 18 months. I’ll give you a four month window there. In under 18 months, you were able to go from you and I walking and having a conversation and looking at a truck to doing a hub to hub demo In less than 18 months.

Xiaodi Hou: In fact, that’s not the case. , we have our first truck by the end of June this year. And before that, we did everything in a simulated environment.

Grayson Brulte: Simulation helped Excel to get to where you’re going to go with your hub to hub demo.

Xiaodi Hou: Yes, especially in the early days, especially in the so called cold start period.

Grayson Brulte: You have the simulation, and how many months has it been since you’ve exported the simulation to put it on a truck to where you began testing? Because that’s obviously a shorter time frame than this overall 14 month period.

Xiaodi Hou: Yes, we are because we’re all crazy. That’s that’s for sure. But in addition to that, the first four months of Bot Auto is that a lot of people, whether you’re in a software infrastructure team or in a motion planning team or control team, or even the deep learning team. Everyone is, focused on building the software infrastructure, and in fact, we have eight people in four months building all of the necessary machine learning infrastructure for the company. Why? Because we use the extensively about what is available out there. We don’t reinvent the wheels. That’s a kind of lightning speed of ramping up with the infra. And after that, we build on the infrastructure, and we certainly, , we built a data so called data collection vehicle. So we started to collect the data, from the wild, from the Houston periphery. And then we’re embracing very wholeheartedly about the pre training process that is, becoming very popular because of the ChatGPT, you know what GPT stands for generative pre trained transformers. So we actually did a lot of pre training techniques in a similar fashion with the GPT. And we use that to accelerate our deep learning code start. And we use transformers, our general structure for deep learning and, months of pre training gives us a lot of advancement in terms of the deep learning capability. And, , based on the simulated environment, we started developing our motion planning and control algorithms. And by the end of June, our first truck arrived at the facility, and we retrofit the truck with, the redundant braking, redundant steering . by, I believe by the end of July this year, we started our, controlled by wire. Basically, the computer is controlling the vehicle first in a closed loop in a private track and later on a shadow test for public highway and now with a human driver, but automated runs on the highway near Houston.

Grayson Brulte: I’m going to put my Wall Street hat on here. Would this be possible without all the breakthroughs that we’ve seen with the GPU architecture over the last 24 months? 

Xiaodi Hou: Not at all. So, you know, I think this, this is really an interesting moment because, , after I left TuSimple, I have nothing, but a bunch of, , strongly minded, similar minded people who also believe in autonomous driving and we wanted to build the company again. Right. And that is the moment where we need, we are forced to start from a clean slate. And we know that this is a basically impossible for us to acquire any of the TuSimple data. So why don’t we just start from scratch? And so we are basically forced to, uh, find the, you know, it’s like a, Spain and Portugal that they are forced to go to the ocean rather than go to the, go to the continent. And they found the brave new world. In a similar way that we have found our brave new world by fully embracing the new era of AI. And I think a lot of things that we’re using today are not available until 2022.

Grayson Brulte: There’s the old saying in life, life is about timing, and it seems that you’ve timed this pretty well yourself.

Xiaodi Hou: Yeah, I think so. I would, but I would maybe ask a little bit, add a little bit more on that. That I think every year or every moment that there is opportunity. And I think really that this moment is like I’m putting myself in the mentality where I’m becoming extremely open and extremely, audacious about new opportunities. And that is what we get in the end. So maybe we can call it entrepreneurship.

Grayson Brulte: There’s entrepreneurship, there’s learn ship, there’s build ship . there’s ship ship, there’s profit ship. We can go down the line on that Xiaodi. Thank you so much for coming on The Road to Autonomy today, clarifying, TuSimple, sharing your vision for bot auto. As we look to wrap up today’s conversation, what would you like our listeners and viewers to take away with them today? 

Xiaodi Hou: There are a lot of ship, ship, as you said, but, , I, think the most important thing, , is courage. I wanted to, , close this with another quote from the Lord of the Ring, which is my favorite book. There is something Good in this world, Mr. Frodo. That is worth fighting for.

Grayson Brulte: You just have to have the courage. If you have the courage, the will, the desire, you will succeed. The Future is Bright. The Future is Autonomous. The future is Bot Auto. Xiaodi,, thank you so much for coming on The Road to Autonomy today.

Xiaodi Hou: Thank you very much.