Two years ago, Xiaodi Hou was ousted as CEO of TuSimple, the self-driving trucking startup he co-founded, which subsequently ceased operations in the United States and was delisted from stock exchanges. In a strange turn of events, the company pivoted in the following direction. Chinese AI animation and games.
Now, Hou is back at a Houston-based self-driving truck startup called Bot Auto. The startup recently made its public debut armed with $20 million in pre-Series A funding. In a sense, Hou is trying to reunite the AV band. Most of his 40-person team is made up of former TuSimple engineers who were laid off during several rounds of layoffs that hit the U.S. team.
Although much of the mission and team are the same, Hou argues that the underlying technology architecture is different. According to its founders, Bot Auto is in the second wave of self-driving car technology, and advances in AI over the past few years have made it easier and cheaper to accelerate AV development.
“We basically developed a machine learning infrastructure in four months with eight people,” Hou told TechCrunch. “In my previous life, it probably took over a year and over 100 people to develop.”
AV2.0 era
Hou isn't the only founder talking about the AV 2.0 era. The AV 2.0 era is attracting new investor interest, despite the collapse of numerous self-driving startups.
TechCrunch recently reported on self-driving delivery startup Nuro's move to a more streamlined AI architecture. Self-driving trucking startup Waabi, which raised $200 million this year, is also touting more advanced AI models. And Wayve, which raised $1.05 billion this year, is developing a self-learning system.
One of the AI advances that helped AV companies in this second wave, including Bot Auto, is Transformers. Transformers are the same architecture behind large language models like OpenAI's ChatGPT. Using Transformers, Bot Auto's large neural network ingests data from a variety of sensors, including cameras, radar, and lidar, to create a “world model” of the vehicle's surroundings, while simultaneously identifying objects and traffic lights. He said he will be able to recognize and plan his route.
Neural networks, which Hou calls “base-to-all” models, stop short of making decisions. It's the data decision-making-output-like flow that occurs in an end-to-end model, which Hou says causes a black box effect. Instead, decisions will be made through “principles-based systems,” which Hou says will lead to “explainable AI.”
Hou said Bot Auto speeds up the process of data collection and processing through pre-training, where neural networks are trained on large datasets before being fine-tuned for specific tasks and environments. You can also.
“Pre-training is key to fast-forwarding training and jump-starting deep learning efforts without having to accumulate years of data and label it all by hand,” Hou said. states. He noted that enhanced supervised self-learning allows Bot Auto to pre-train the model from unlabeled data without human feedback, significantly speeding up the process.
Hou said Bot Auto allows him to take advantage of open source tools that weren't available when he led TuSimple, so he doesn't have to think about how to tune GPUs or perform high-performance data storage. I did.
For example, Hou said Bot Auto uses Kubernetes as its cloud operating system to automate nearly all critical data-related tasks, including large-scale simulations and mining road test data.
Focus on operations
Bot Auto is already testing self-driving trucks on Texas highways. Image credit: Bot Auto
According to Hou, eliminating the critical task of building AV infrastructure means Bot Auto can focus more on its core business of bringing AVs to market.
“Of course, we always try to keep up with the best technology, but we're not here to show off technology,” Hou said. “We're here to facilitate the work and make it happen. That's why the operation is in Houston.”
Texas is a leader in self-driving trucking, and not just because the state's sunny weather makes testing easier. Texas is home to some of the most profitable freight routes in the country, including Interstate 45, which connects Houston and Dallas. Both Waabi and Kodiak Robotics test trucks on these routes daily.
And now, so does Bot Auto.
Hou said his startup has begun testing self-driving trucks with human safety operators in the front seats on public roads in Texas, and has already secured two letters of intent from customers in the logistics and shipping industry. He said there was. Bot Auto is also on the right track to complete a hub-to-hub demo in 2025, Hou said.
Despite the scalability of its base model, Bot Auto is cautious when it comes to scaling up its ground operations.
“We're going to contain ourselves in a very cost-effective way,” Hou said, noting that investors in 2024 are only investing in sustainable and genuine technologies. “Our operating fleet is very small. We are focused on a small number of lanes and will only begin to scale up once we know how to save costs and operate efficiently.”
Not without challenges
Despite the win, Bot Auto may face headwinds.
Investors may be warming to the idea of a new era for AVs, but they haven't forgotten the challenges of the past few years, including the closures of Embark Trucks and Waymo's self-driving cargo operations.
And, of course, there were the failures of TuSimple, the worst of which came after Hou was ousted as leader.
TuSimple's board had voted to fire Hou after an internal investigation revealed some overlap between the company and Chinese startup Hydron. Hydron was founded by Hou's co-founder Mo Chen in 2021 to produce autonomous hydrogen trucks, around the time TuSimple went public in the US.
Although Hou claims he was fired without cause, such a public firing, and one made worse by geopolitical tensions, could give investors pause. Company officials have backed up that claim to TechCrunch in the past. Hou also said that experience has taught him not to try to cross the line in unfriendly countries, which is why Bot Auto is entirely a U.S.-based company and has no ties to China. He said this is one of the reasons.
Asked why he didn't launch Bot Auto in China, where regulations are more favorable to AV companies, he said the high labor costs for truck drivers in the U.S. increase the potential value of autonomous driving. Ta.
“It was a business decision,” he said.
Additionally, there is also the threat of legal action from TuSimple itself. Hou retained his seat on TuSimple's board after being fired, but eventually resigned in March 2023 after being accused by the board of trying to poach staff for new ventures. Hou denied this at the time, telling TechCrunch that he only announced his decision to form a new company after stepping down as director. He said the TuSimple alumni who now make up his staff came to him out of loyalty and a desire to continue working in AV after being laid off.
It is unclear whether these accusations will resurface. TuSimple CEO Cheng Lu told TechCrunch that the company will likely sue Hou for intellectual property and trade secret theft.
Hou said he was not worried.
“Everyone here is suing everyone,” he said. “And after talking about the technology that we're building, I don't think much of what we're doing would have been possible before 2021.”