TuSimple and Waymo lead the way in the emerging autonomous freight sector; TuSimple founder Xiaodi Hou and Waymo truck driver Boris Sofman had an in-depth discussion of their industry and the technology they are building at TC Mobility 2020. Interestingly, although they solve the same problems, they have very different backgrounds and approaches. .
Hou and Sofman started talking about why they pursued the truck market in the first place. (Quotes have been edited slightly for clarity.)
“The market is huge; I think in the United States $ 700 – $ 800 billion a year is spent on the trucking industry. It continues to grow every year,” said Sofman, who joined Waymo from Anki last year to expand its freight efforts. . “And there is a huge shortage of drivers these days, which will only increase in the time to come. It’s just such a clear need. But it won’t be overnight – there are still a whole host of challenges. You can’t avoid it. So the way we talk about it is that the things that are the hardest are just different. “
“It’s really the cost and reward analysis when you think about building the operating system,” said Hou. “The cost is the number of features you develop, and the reward is basically the number of miles you drive – you calculate per mile. That cost-reward analysis shows that freight is just the natural way for us to go. total number of problems to solve is probably 10 times less, but maybe, you know, five times as difficult. “
“It’s really hard to quantify those numbers,” he concluded, “but you know what I mean.”
The two also discussed the complexities of creating a perceptual framework that is good enough to drive.
“Even if you have perfect knowledge of the world, you have to predict what other objects and agents in that environment are going to do, and then make up your own mind and the combination knows it’s very challenging,” Sofman said.
“What really helped us is the realization on the automotive side of the business many, many years ago that we needed to make our own sensors to help us solve this problem as easily as possible and ease the challenges downstream,” he went on. “And so we have our own lidar, our own radar, our own cameras, and they have incredibly unique features that are custom designed through five generations of hardware that try to really lean into the most challenging of situations that you just can’t avoid on the road. “
Hou explained that while many autonomous systems are descended from the approaches used 15 years ago in the famous DARPA Grand Challenge, TuSimple’s is slightly more anthropomorphic.
“I think I’ve been deeply influenced by my background, which has a touch of neuroscience. So I’m always thinking about building a machine that can see and think like humans do,” he said. “In the DARPA challenge, people’s idea would be, Okay, write a dynamic system equation and solve this equation. For me, I’m trying to answer the question, How do we reconstruct the world? That’s more about understanding the objects. , understand their attributes, even though some of the attributes may not directly contribute to the entire self-driving system. “
“We are combining all the different, seemingly useless features together so that we can reconstruct the so-called ‘qualia’ of the perception of the world,” continues Hou. “By doing that, we discover that we have all the ingredients we need to complete any missions.”
The two disagreed on the idea that because of the wide differences between highway driving and street-level driving, there are essentially two different issues that need to be resolved.
Hou felt that “the overlap is quite small. Human society has established certain kinds of rules for highway driving … this is a much more regulated system. But for local driving, there are really no rules for interaction. In fact, very different implicit social constructs to float in different parts of the world. These are things that are very difficult to model. “
Sofman, on the other hand, believed that while the problems are different, solving one substantially contributes to solving the other: “If you break down the problem into the many, many building blocks of an AV system, there is a quite big leverage where even if you don’t solve the problem 100% take away 85% -90% of the complexity We use the exact same sensors, the exact same computing infrastructures, simulation framework, the perception system is largely carried over even if we re-order some models The core of all our algorithms is: we are working to keep them the same. “
You can see the rest of that final exchange in the video above. This panel and many more from TC Sessions: Mobility 2020 are available here to view for Extra Crunch subscribers.