Tony Han Wants a Self-driving Car on his Grave.
When asked about his plan for his 3-year-old startup WeRide to go IPO, Tony Han (Chinese name Han Xu) simply answered “IPO is like the wedding. Whether the marriage afterwards is harmonious and great has absolutely no correlation to the wedding, which is merely a celebration in itself. I am working day and night to make our technology happen, and I just wish to have a self-driving car on my grave when I die.”
Below is the full adaptation of the original Chinese interview of JazzYear (a Chinese Tech Media company) with WeRide’s CEO Tony Han, which took place on July 30, 2020.
Interview Topic: Can we see the real fully self-driving cars on the streets in 5 years?
Interviewer: The fully open operation of autonomous driving is an extremely daunting attempt in the entire industry and the world. Our audience is very interested in the progress of your trial Robotaxi operation in Guangzhou. Can you expand on some of the difficulties in the operation during the past six months?
Tony Han: To be honest, the biggest difficulty has been the Covid-19 situation. People are always more difficult to manage than machines. It is a big challenge to manage the safety aspect while trying to maintain the normal operations under the influence of the pandemic. During the epidemic, not only did the number of passengers decrease, but we also needed to spend a lot of time and energy on the necessary precaution, including training our “safety drivers” to wear masks, disinfection, temperature measurement, and signing on duty. However, all these elements also gave us the opportunity to improve our operations and management. As the largest state-owned taxi company in Guangzhou, our partner Baiyun Taxi Group has very rich management experience in both taxi drivers and safety drivers, which has helped us a lot in navigating this difficult time.
Interviewer: Your Robotaxi operation adopts the “iron triangle” model, involving the partnership between the Autonomous Technology Solution provider (WeRide), Automotive manufacturer OEMs and a ride-hailing platform. How did the various companies participate in the Guangzhou on-the-ground operation? What is the key logics behind the business model and the capital flow?
Tony Han: In the “Iron Triangle”, Baiyun Taxi Group is mainly responsible for operation and management, Science City Group is responsible for coordinating government relations and providing fund support, and WeRide is responsible for providing the technology. Everyone contributes their respective strength to the partnership.
Interviewer: Do I have to pay if I order a WeRide Robotaxi car in Guangzhou?
Tony Han: If you use the “WeRide Go” app, it will follow the billing standards of ordinary taxis.
Interviewer: Do you have any specific feedback from the more than 4,000 users who have taken Robotaxi so far?
Tony Han: There are many feedbacks. I have read the feedback report. Some users were dissatisfied with the fact that our arrival time is not accurate. From this point of view, there is still a gap between us and a company that does a pure ride-hailing matching operation. The amount of data we have is limited due to the fact the Robotaxis can be ordered directly on our app or through a third-party app such as AutoNavi Map. For time estimation, we need to rely on third-party platforms and therefore has its limitation in terms of accuracy.
Another observation is the question of whether riders should look for the car, or the car should be looking for the rider. When using the common ride-hailing app (say Didi), if the driver cannot find you, he will call you and say “wait for two minutes” so the rider does not have to move. When it comes to self-driving Robotaxi, we can accurately locate the passenger’s location within centimeters on the map. But if passengers want to ask where the car is (in the near future when all Robotaxi will not have any safety drivers), there is no way for the riders to ask. Therefore, the public has to get used to not having to call at all, because most people still think that the map can be off by a few meters sometimes.
The third point being that some passengers felt that the safety drivers were acting cold (completely ignoring the riders). This is actually an important rule that we have, in order to allow the safety driver to maintain the highest alertness on the road condition. The main responsibility of the safety driver is safety. But some passengers hoped that the safety driver can also serve as a guide to explain how the autonomous technology works during the ride. These are the main problems we discovered during our operations, and we are currently tackling them one by one.
Interviewer: How does WeRide evaluate the advantages and disadvantages of Didi’s competition?
Tony Han: Didi is a great company, and its greatness lies in solving the problem of matching ride-hailing services through the Internet. The combination of autonomous driving and the Didi platform should be an optimization because a platform like Didi should manage a range of mobility needs.
So in the long run, I hope that we can have a cooperative relationship with Didi. I hope that everyone will work hard together towards the goal of achieving this important fully autonomous technology. Everyone should uphold an attitude of tolerance, fairness, and cooperation. When your competitor have done something great, instead of being angry, we should be applauding to them.
In fact, our relationship with Didi has been very good. I personally am very good friends with Didi CTO Zhang Bo and Didi Autopilot COO Meng Xing. (I just saw Meng Xing in the another live broadcast room). We have regular communication and I certainly hope to deepen our official cooperation in the future. Everyone is working to create their own version of the self-driving technology now. But when one version is “almost perfect”, that will become the center of the attention which will shift the competitive dynamics of the industry.
Interviewer: Autonomous driving has always been controversial, as many argue about the technology trends as they emerge. For example, a few years ago, many discussions were centered around the choice of technical routes: Is it a step-by-step approach from L2 and L3, or a one-step approach such as L4 and L5? Should we only use cameras, or we need the expensive Lidar sensors? In terms of the technology platform, should ie be closed, semi-open, or completely open. Today, what do you think about the prospect of these paths ? What are your own selection criteria?
Tony Han: In 2017 and the first half of 2018, there was a fierce debate about whether to use a pure cameras or cameras plus Lidar sensors. The key question was whether to adopt Lidar for autonomous driving, as the cost was higher. At that time, I made it really clear that Lidar are absolutely necessary, because personally I have an extensive background in computer vision. I knew that relying solely on camera vision is just not sufficient to achieve the extremely high level of stability of our algorithm, from the safety perspective.
Now. no one is arguing about this topic, basically all of them are using Lidar.
Regarding the dispute over the L4 path, I firmly believe that the path from L3 to L4 just does not work. L3 level vehicles can assist the driving on specific conditions such as on high speeds. In all the accidents with L3 vehicles, the safety drivers all became liable for the legal consequences. This creates a fundamental paradox, in which the responsibilities of human versus machine are not allocated fairly.
Trying to achieve L4 through a transition from L3 is like trying to solve a problem using arithmetic instead of using calculus, which is what is actually required. For example, we can never calculate the surface area of a sphere if we do not understand the limit of the arithmetic approach which is simply using addition, subtraction, multiplication and division.
If you want to achieve L4, some problems must be deeply thought through and resolved from the very beginning of the process. We can never successfully transition from L3 to L4, and that is a fact. Another example is that L4 autonomous driving absolutely requires the use of Lidar sensors and high-precision maps. However, many L3 car manufacturers did not use them due to cost and coverage considerations. These are the type of complex problems that cannot be solved by simply tweaking our algorithm.
Interviewer: Should autonomous driving be based on “single-vehicle intelligence” or “vehicle-road collaboration”? The key difference between how China and the United States approach the development of the autonomous technology is that most U.S. companies focus their efforts on building AI intelligence based on a “single vehicle”. In China, due to the national policy and support, the focus seems to be transforming the roads and cars at the same time. I want to hear the views of you and your company in this regard. What do you think about the end of China’s autonomous driving future?
Tony Han: This is a technical discussion. Let me try to answer as an engineer, and not the CEO of WeRide. From a technical point of view, in the future the car and the road must be coordinated, because both exist to serve the single goal of mobility. However the vehicle-road collaboration involves the proper division of tasks between the vehicle and the road. And this is a big challenge that requires a lot of brain power.
Every one seems to be rushing to do the things that are obvious and easy when it comes to the V2X (Vehicle-to-X) infrastructure. But no one wants to do the difficult things. From my point of view, it is necessary to let the car solve the “common problems” and the road to solve “the other problems”. People usually think that the road is very stable, but it is not. I have built a small mathematical model from the perspective of practical vehicle-road collaboration, and I will share with you.
In the vehicle-road collaboration, a pole is installed every 200 meters on the road. A set of sensor modules are installed on the pole, including 1 or 2 Lidars, 2 or 3 high-definition cameras that take care of the left and right sides, and a host on the edge. This system usually restarts every few days or a month, as our personal computers might. Now let’s look at the stability of this system. Let’s say the system operates normally in 364 days out of 365 days in a year without restarting, the stability level is 99.7% (364÷365). When you travel 20 kilometers, you need to pass 100 poles. If our autonomous driving completely relies on vehicle-road coordination, none of these 100 poles can be broken. What is 99.7% to the power of 100? It is 0.76. In other words, if you rely solely on vehicle-road coordination, on a 20-kilometer road, there will be 1 out of 5 times (1–0.76=24%) that your car will be stopped because of the unavailability of the system. Now, do you dare to sit in such an autonomous car?
When we use a mathematical model to quantify this matter, we will find that autonomous driving requires both the “single-vehicle intelligence” as well as the “vehicle-road coordination” to work together. It is impossible to realize autonomous driving by putting all the technologies at the end of the road and using a vehicle as the receiving end, from the model I just illustrated. This is why we are going full speed on building the “single-vehicle intelligence” , while also thinking long and hard about the relationship between the car and the road.
Many technical elements require in-depth thinking, hands-on implementation and iterations. For example, how do we arrange the sensors? How stable are the sensors as a whole? We prefer to keep our heads down on solving the hard problems, instead of just chanting marketing slogans (before we have any real progress). If all we need are just some sensors on the road to achieve full autonomous, the Berkeley research project that started in 1965 could have brought us the results we want. (The research project was conducted by University of Berkeley with the National Automated Highway System Consortium (NAHSC), leading to the historical “Demo ‘97” demonstration of platooning of eight AVs guided by magnets embedded in the highway and coordinated with vehicle-to- vehicle (V2V) communication.)
Interviewer: What do you think about the choice of these two paths, relying entirely on “single-vehicle intelligence” and “vehicle-road coordination”?
Tony Han: It depends. Waymo has been working on their “single-vehicle intelligence” for more than ten years now, but everyone’ patience seems to be running out.
Interviewer: It is very difficult for the United States to rebuild roads, so they put all their energy on rebuilding the cars. What do you think of their choice? According to your understanding, will China and the United States take two completely different roads in the direction of autonomous driving?
Tony Han: I don’t think we will go a completely different path. Everyone is groping and no one is really doing well now. If one way can be proven, the other ways will just quickly converge. If you look at all the leading autonomous driving companies, whether it be Cruise, Waymo, Baidu or WeRide, the layout of most sensors is not fundamentally different. Some things will quickly converge through practice, but the perfect answer has not been found yet, so it is still a debate in the air. We used to ask the question: what should the perfect smartphone look like? No matter it is Android or iPhone, now we know it looks like this (taking out his phone). When something is perfect, no one needs to argue anymore.
Interviewer: In the first half of this year, the pandemic had brought many industries to a crisis. Now looking at the tremendous “new infrastructure” plan that the government is pushing, with autonomous driving being an important part of it, what kind of impact does such a positive force have on your business development and your layout in the vehicle-road collaboration?
Tony Han: This is a wonderful thing. We are also very excited about the “new infrastructure” opportunities. This is why we are cooperating with Gosuncn, China Unicom, and AutoNavi, and we are also seeking cooperation with SenseTime.
There are not many opportunities to define the future, but many are indeed in the “new infrastructure”. The resources of our society are limited and the limited resources need to be used well. Therefore, we hope to actively participate in the new infrastructure construction, in order to solve the problems of traffic safety and efficiency. We will also actively participate in bidding and technical discussions with local governments. Even if the vehicle-road collaboration is not for L4, but for L2 and L3, we can also contribute some of our accumulated technology and experience.
Interviewer: When we talked about autonomous driving and its future benefits, it was actually an iteration and improvement of the original taxi model. What you really end up serving is the users, which means the income will come from their wallets. However, under the wave of new infrastructure , companies like yours will expect significant revenues from the government’s commercial projects. What do you think about this China’s way of achieving the common goal?
Tony Han: If the government does not provide infrastructure such as water and electricity, it is actually very difficult to achieve our goal. Road construction is basically infrastructure construction. The government should take the lead and bring experts in various enterprises to come up with truly practical and efficient solutions. The idea is to spend the least amount of money and getting the most bang for the bucks.
However, before building more new infrastructure, a few things need to be in place, namely thorough argumentation, clear planning, as well as national standards. We have seen that during the construction of poorly managed autonomous infrastructure, some roads were dug up, then rods and sensors were buried, only shortly afterwards to be removed and redone again. This would be a horrible thing because of the waste of material and productivity.
Interviewer: If the full score of achieving full autonomous driving is 100 points, how many points would you give yourself now? Are there any technical gaps?
Tony Han: This question is like an English saying — — “unknown unknowns”. There are some things we don’t know we don’t know. Sometimes I think it accounts for 30% to 40%, but in fact there is still a lot of uncertainty about the remaining 60% to 70%.
I think we now have only 30 points to achieve perfect autopilot, but we have achieved 70 points for our commercial operations. I have always believed that we must not wait for perfection to operate commercially. Even though we still have quite a long way ahead, we must get the wheels turning under the premise that safety is ensured. If you look at our human history, things like cars, airplanes, and wireless communications have been quickly introduced to the market when they can be used safely. Then the natural competition in the market will trigger capital investment and innovations so the technology gradually ramped up to maturity.
Interviewer: I want to ask a more realistic question. As you said, during the entire six months of operation, there were zero accidents and the overall safety level was high. But, how would you handle any accidents if they do happen in the future?
Tony Han: Autopilot is still in the testing stage. There is a safety driver in the driver seat in every car. You can look at it as an ordinary car on the current ride-hailing network. Now, the safety driver has chosen to “turn on” the autonomous driving function. so he or she will be responsible for the safety of the ride. Because of this principle, every single safety driver must undergo very rigorous training and certification before deploying to our operation.
From a mathematical point of view, no one can prove the absolute stability of a software system. Any operating system we use today may crash. But from the current statistics, self-driving cars indeed are safer than human drivers, because it is difficult for human to drive 2.6 million kilometers without accident (that is the total mileage our fleets have accumulated with zero accidents).
In the future when full autonomous driving is realized, who should be liable for any collision? I have always appealed to the outside world, hoping that L4 level autonomous driving can adopt the “elevator theory”: No one in the world can prove the absolute safety of an elevator. If it has a problem and causes injury, the elevator manufacturer will be held accountable because it is deemed as a product quality issue. We have had many discussions with legal experts before, and as it stands right now, the safety liability of autonomous vechicles most likely will follow a similar framework.
Interviewer: You (WeRide) have obtained the Remote Control Autonomous Test Permit from the Guangzhou Government. Therefore you are now allowed to remove the safety driver in your Robotaxi in restricted areas. Will WeRide have plans and preparations for fully unmanned testing next year?
Tony Han: It will depend on the progress of our technology. But I want to emphasize that safety is always our priority. It is necessary to ensure not only the safety of the software system after complete development, but also that during the development process. We plan to conduct tests in some semi-enclosed venues in the near future to ensure the stability of fully driverless (no safety drivers) conditions. When our technology is fully tested, we will continue to apply to the relevant departments for further testing of fully driverless (without safety drivers) in open roads. In addition, we will also apply and test fully driverless (no safety drivers) in the United States to get to the next commercial operation milestone there as well.
Fully driverless is a new thing and must be promoted in policies and regulations. Just like before the birth of the airplane, no one knew that people could fly in the sky. But if they didn’t try and perfect the technology along the way of implementation, humans would only be able to stay on the surface of the earth forever. The same is true when it comes to fully driverless technology, which is something that no one has ever done before.
Interviewer: Your work may involve dealing with traffic and other regulatory agencies. From the perspective of legal frameworks, how many more years are we away from being able to remove safety drivers completely on the open road?
Tony Han: At present, I think no one can answer this question. But in recent years, we have witnessed the fully driverless technology being able to operate with very high stability in a relatively simple area. My view has always been this: Assuming that a person’s effective driving period is 20 years (or the equivalent of non-stop driving mileage). If the autonomous system can drive for longer than that without once the need for human takeover, then we will have the confidence to remove the safety driver from the car.
In addition, fully autonomous is not completely unmanned, because someone will be monitoring the status of the car online all the time. Don’t worry too much about the remote control problems. Do you really think in our cars now, the pedal that you press down is directly connected to the brake? Do you really think when you turn the steering wheel, you are actually turning the wheels directly through mechanical devices?
In fact, many vehicles’ brakes and steering wheels are already controlled mechanically through electrical signals. So remote controlling the car simply means putting that digital connection through the internet. In terms of exactly when this can be achieved will depend on two factors: 1) when the technology and regulations are sufficiently mature; and 2) when people can generally overcome the psychological barrier of the remote-control concept. Any progress in this matter requires not just boldness, but a tremendous amount of cautions. Take landing on the moon as an example. No one could ever deny the danger of such concept. But when we made the technology stable and reliable enough, we would be able to say okay let’s take the big step forward. The danger (i.e. risks) of landing on the moon did not prevent humans from trying and succeeding.
Interviewer: This may be the fate of any practitioners in the space. Before the arrive of that day, nobody can stop reasoning and persuading for this grand vision becoming a reality.
There was a question from an audience in the live broadcast room: “Hello Mr. Han, I had the opportunity to experience your Robotaxi in Guangzhou. Do you have any requirements for the next round of investors? Is there a timeline for an IPO for the company?”
Tony Han: An IPO timeline is difficult to predict. It is more like a wedding ceremony, which in the end has nothing to do with whether the final marriage is a happy one. When I started the business, I realized that IPO is just a bell ringing ceremony. It is not a particularly important thing. After going public, the company still needs to face more pressure. So it means that the company itself doesn’t become particularly great simply after an IPO. In fact, companies that are not great might end up delisting due to undervalued stock price.
If my purpose of starting a business is simply an IPO and converting my shares to cash, I probably might not have even done it. An IPO in itself does not really have much meaning to me. In terms of requirements for investors, I hope that they can support us in building the autonomous driving technology, which can be the ultimate platform that allows people to achieve 4 things in mobility: safer, more economical, comfortable and convenience.
When I leave this world, I hope that my tombstone is engraved with a self-driving car. As for when our company will be IPO, I don’t think it matters at all.
Entrepreneurship is running 30,000 meter race
Interviewer: Just now we talked a lot about policies, technology, strategy and the current status of the industry. Let’s switch gear to your personal life. I would like to ask you: as the founder and CEO of a growing technology company, what are some of your new personal realisations in 2020? Anything important that you learned recently that were not obvious when you started the business?
Tony Han: I enjoy reading, and I always recommend books to friends around me. There is a book that I would recommend to any entrepreneur before starting a business: “The Hard Thing About Hard Things”. The Chinese name is “Chuang Ye Wei Nan”. Entrepreneurship is like running a long-distance race. When you reach the 300 meter mark and start thinking how far the finish line might be, you typically find out that you are actually in a 30,000-meter race. To me, it is all about perseverance of the mind and the body. It is hard but at the same time very interesting.
Interviewer: In the next three years, what do you envision your company WeRide could achieve? Are you satisfied with whatever you have accomplished so far?
Tony Han: To get to this date, I have already given it at least 100% of what I have got. So I’d say I am quite satisfied with our progress. But as we continue to push forward, the goal will inevitably be raised on a continuous basis. In fact, I look at myself as a new me every day, and I do let my goal adjust itself. As Thomas Carly said “cease to struggle, cease to live.”