With offices in Beijing and San Diego, TuSimple LLC is developing a full stack software solution for autonomous trucks using computer vision algorithms, millimeter radars and 3D HD mapping technology. In an interview with automotiveIT, Co-Founder & CTO Xiaodi Hou explains why autonomous trucks make more sense than self-driving cars and talks about the challenges of being a hybrid US-Chinese company amid rising trade tensions. And he warns that accidents caused by sloppy coding could jeopardize the whole autonomous vehicle industry.
Why autonomous driving trucks?
I’m really good at making algorithms and I’m really paranoid about trying to make a perfect system. So [I thought that I would] let my new venture reflect the value of this tendency to require perfection. You could say autonomous driving is a very obvious choice because the more reliable the system you create, the more value it has. So initially we started talking about doing ADAS or driving-assist systems. But in China you have a lot of different road conditions. Sometimes there is no road at all. If you were a foreigner in Beijing you wouldn’t even know how to drive. There’s a lot of implicit understanding of human society needed in order to be able to drive. Knowing that fact we began thinking that, maybe, autonomous urban driving isn’t a solid idea. And one day we had the idea to focus on freight transportation on the highway.
Nvidia has made a strategic investment in TuSimple. Who else has been funding you?
The primary backers in terms of total money invested are a composite fund from Hong Kong and Sina Corp. My previous company had a pretty strong collaboration with Sina. There’s a lot of trust between Sina and our technology and the whole TuSimple team. So, the first funding round was by Sina. The second round was led by Sina, and the third round is led by this composite fund.
How many miles have you tested so far?
We have a testing facility in Tucson and in Beijing. We’ve done a little more than 15,000 miles. But I’m not very particular about this number and we’re not eagerly accumulating miles. We test on a checklist basis. We have a checklist on what is working and what is not working perfectly. We are dedicated to testing those things that are not working.
Did the recent accident with an Uber car change your testing program?
This was a big thing and we had to stop our testing for a period of time. This was kind of in response to (Arizona) governor Dusey not wanting to intensify the situation. Also, Jensen Huang, CEO of our very important partner Nvidia, said all companies should pause and check their systems. We still have our vehicles collecting data, but we are limiting our testing. We have two guys on the vehicle, and after the Uber accident we made this a complete 100 percent rule. We are more fearful of an accident than any government is because it would kill the entire company.
Tell us about your safety systems.
We have an insane level of redundancy and safety requirements. For example, in an area such as localization, one algorithm would give you 99.9% of accuracy, but that’s not enough. We have a second algorithm, so that when the first one fails, the second is still going on. That gives you 99.9% accuracy or reliability. And then even if the second one fails, we have a third algorithm that is compensating. The three algorithms will always work together, cross-checking each other. The other part of the effort is that we try to offload some of the computation to an off-line module. This is philosophical; we do not trust the on-line module. So, whenever we have a chance to do the computation off-line, for example, we do the map off-line.
You don’t use lidar, preferring to use computer vision technology because of its longer range. Could that change?
I think there’s a big misconception that I naturally hate lidar because I’m computer-vision based. At TuSimple, we welcome all of the technologies. We have tested all sorts of weird sensors that you have never even heard of. One issue with lidar is that the commercial viability is just not there yet. We have a lidar team and we create our own lidar algorithms. However, even if we know that we’re doing better than the industrial standard we’re still not yet at the level of our camera perception algorithms. Adding lidar is not adding any value. However, when the cost of lidar starts going down and the performance is going up we could think about using lidar in the future. We’re totally open to that.
Who do you see as your primary competition? And what is the biggest risk you run?
Our biggest competitor is time. I’m very confident that the code that we have is very rock-solid. In many of the companies we see around us the code is like a bad imitation, like the poorest region of a wealthy country. There are some academic people who want to prove that you will always kill someone in certain scenarios (involving autonomous vehicles). But that is completely useless. What really kills people are accidents caused by bad implementation.
You have headquarters in Beijing and San Diego. What is it like to be a hybrid US-Chinese company?
I am literally working until four, five A.M. every day. But, we progress very fast that way. We are also in a good position to attract talent both in the US and China. We can do fund raising in both countries. Those are the good parts of it. The problem is that we are working overtime to communicate. I have to make an extra effort to achieve smooth communication. But I think things get done faster with people in both places.
Could the trade dispute between the US and China impact your operation?
I don’t think so. The trade barriers are already there. Chinese trucks are extremely cheap, but the US government, with the help of emission regulations, is making sure that no Chinese truck will ever enter the US market. On the other hand, US trucks are extremely expensive and have no competitive power to enter the Chinese market.
When will you be able to roll out your product for actual commercial use?
I feel the technical stack will be almost stabilized by the beginning of next year. However, we do not want to roll out a real product until the end of next year. We still need a lot of testing, a lot of validation to prove the technology to ourselves and to the public.
Interview by Jason Booth