A Conversation with PENG Jun from Pony.ai: A "Counterintuitive" Autonomous Driving Company

In 2016, when Peng Jun founded Pony.ai, he already anticipated that it would take at least a decade of effort for Robotaxi (autonomous taxis) to transition from a vision to reality and achieve large-scale implementation. This would require the maturation of technology and laws and regulations, as well as consideration of social acceptance.

Today, whether it's Waymo shuttling through the streets of San Francisco or Pony.ai's fleet taking orders in the bustling downtown of Nanshan, Shenzhen, it confirms this point. Autonomous vehicles have already integrated into the social traffic flow and started commercial operations.

However, what Peng Jun might not have expected a decade ago was that a key issue restricting the rapid expansion of the fleet is a series of trivial maintenance tasks after the absence of drivers.

When humans drive, tasks such as charging, car washing, vehicle maintenance, and even helping customers carry luggage are all done by the driver. However, when autonomous vehicles hit the road, these seemingly insignificant tasks have become problems.

Peng Jun, the CEO of Pony.ai, recently participated in an interview with 36Kr Auto. He discussed topics such as the implementation and operation of Robotaxi, the choice of L2 mass production business, why there is no rush to layout embodied intelligence, and the selection of technical routes.

Peng Jun told 36Kr that Pony.ai has established an operation team and explored a set of maintenance standards to provide logistical support for autonomous vehicles. "These personnel include both remote safety operators and on - site support and logistics staff."

In the future, even if the fleet size expands rapidly, the vehicle - to - personnel ratio will not increase significantly (that is, breaking the traditional growth logic of more vehicles requiring more people). From this perspective, Peng Jun believes that this is a blind spot in capabilities that ride - hailing companies or automobile manufacturers may overlook when entering the Robotaxi field.

After a decade - long journey, along with the rapid maturation of the supply chain driven by automotive intelligence, unmanned vehicle technology has entered the stage of substantial commercial operation in various fields. The representative example is unmanned logistics vehicles. Companies such as White Rhino, New Stone Age, and Jiushi Intelligence have successively received capital investment and have become the business card of China's automated logistics.

As one of the markets with the greatest potential for unmanned vehicles, Robotaxi has attracted automobile manufacturers, ride - hailing platforms, and other companies to enter the market one after another. Tesla, XPeng Motors, Geely, etc. have all announced clear unmanned vehicle operation plans.

A mainstream view in the industry holds that automobile manufacturers, with their complete vehicle engineering capabilities and the technical system from assisted driving to autonomous driving that can create a data snowball effect, will be the core players in Robotaxi. However, Peng Jun told 36Kr that manufacturing capabilities have overflowed in China. For Robotaxi, if a company has never been involved in it before, it is almost starting from scratch.

Peng Jun believes that for Robotaxi, technical capabilities determine the transition from 0 to 1, that is, whether the company can undertake this task, while operational capabilities determine efficiency. Currently, Pony.ai has gradually developed a clear business model through exploration.

Throughout the development process, the path of Robotaxi has not been smooth. It has long faced commercialization challenges. Many peer autonomous driving companies have shifted their focus to the L2 intelligent driving supplier business to find more revenue channels.

Pony.ai is one of the few companies that has consistently adhered to the Robotaxi front. Currently, Pony.ai's seventh - generation vehicles have achieved single - vehicle profitability in Guangzhou and Shenzhen. After the verification of the financial model, Pony.ai plans to rapidly increase the fleet size to 3,500 vehicles this year.

Looking back on the choices made along the way, Pony.ai has mostly made "counter - intuitive judgments".

Peng Jun said that the company once briefly tried the L2 assisted driving mass production business but quickly found that it was destined to be a low - profit industry. "Due to the low technical threshold and non - standard user experience, automobile manufacturers have the say, and intelligent driving companies are prone to get involved in price wars."

Facing the aggressive layout of Robotaxi by automobile manufacturers and ride - hailing platform companies, Peng Jun sharply commented, "Making announcements is always an easy thing." "Tesla has been talking about it for 10 years. Has it achieved it?"

Automobile manufacturers generally hope to follow Tesla's example and use their existing end - to - end algorithm capabilities as the same - source technology to support both assisted driving and Robotaxi businesses.

However, Peng Jun provided a different technical judgment from the mainstream in the industry during the interview. Pony.ai did not follow the technical rhythm of large language models to build an integrated autonomous driving algorithm with a super - large number of parameters. Instead, it adopted an algorithm strategy with multiple small models to improve operational efficiency and reduce computing power dependence.

In response to the situation in the assisted driving industry where "the mention of maps causes panic", Pony.ai also did not follow the trend. Peng Jun admitted that the company will still adopt the light - map route in the long term.

"You drive more easily on the roads you are familiar with, and you feel more tired in unfamiliar places. This is very normal. So why not use maps?" Peng Jun said. "Even Tesla uses maps."

In the booming field of embodied intelligence, Pony.ai did not rush to enter the market but chose to observe.

"This is another thing that will take at least a decade," Peng Jun gave a more straightforward judgment. "There will always be opportunities to get involved, but it's still necessary to see clearly and think carefully."

Peng Jun, CEO of Pony.ai

The following is the edited transcript of the conversation between 36Kr Auto and Peng Jun of Pony.ai:

On the implementation of Robotaxi: Automobile manufacturers and ride - hailing platforms are not good at it

36Kr Auto: I remember that in the early days, the industry was saying that Robotaxi (autonomous taxis) would be implemented around 2020. However, judging from the progress of Waymo or your company, it seems that this goal has only been achieved today, about five or six years behind schedule. Where do you think the gap lies?

Peng Jun: I think it was the people who didn't understand that made such claims. I've always believed that Robotaxi requires at least a decade of effort.

Just like Elon Musk saying every year that Robotaxi will be realized. He's been saying it for 10 years but hasn't achieved it yet. These are all claims made by people who don't understand because they clearly don't understand what L4 is. Those who were very vocal at the beginning, like Cruise under GM, have now gone bankrupt.

I think those who really understand know the complexity. It takes at least a decade of effort from the development of technology, the maturation of laws and regulations, to people's acceptance.

36Kr Auto: Now that Robotaxi has been put into operation, many problems have been magnified. For example, sometimes Waymo's vehicles stop at traffic lights or drive onto railway tracks. Are these the most difficult parts of implementation?

Peng Jun: These problems are endless. They will always emerge, and we can only address them one by one. In this world, many things are already 99.99% in place. To achieve the last 0.01% or 0.001%, we definitely need to keep working.

36Kr Auto: What kind of mechanism should be used to solve these problems?

Peng Jun: There are several systems. First, just like treating a disease, early detection. After detection, we can quickly incorporate general solutions into the development cycle and add them to the boundary conditions. The world model we've built can accurately model the surrounding environment of the vehicle, including accurately representing the kinematic model of the vehicle itself and the kinematic models of surrounding traffic participants. By continuously enriching the scenario data in the world model, when the system detects new special situations, it can quickly include them in the training samples of the world model, enabling the model to learn to handle such boundary conditions and thereby improving the generalization ability of the autonomous driving system.

Moreover, in such cases, a good backup condition must be set.

36Kr Auto: What is your backup mechanism like?

Peng Jun: There are many. First, there is technical backup. For example, we have a lot of redundancy in the design. All sensors are redundant. The entire vehicle control, including acceleration, braking, steering, power, electricity, and network, is self - redundant.

Secondly, we have a whole set of detection and fail - over mechanisms. You can imagine it as a three - level degradation mechanism. Our main system is for normal driving. Usually, 99.9% of the time, it's the main system. If the main system really malfunctions, we actually have a mechanism to pull over. If it's on the highway, it can even pull over and find the nearest exit. So, it's a degradation mechanism.

If this level also fails, in the worst - case scenario, the vehicle will stop safely within the lane lines. It still has sensors, but the third - level situation is not ideal as it will block the road.

However, there is also remote monitoring, which will detect problems in a timely manner and summon ground support personnel to the scene for assistance.

36Kr Auto: You've set a goal of 3,500 vehicles this year. How was this figure calculated? Why 3,000 instead of 5,000 or 10,000?

Peng Jun: Production is relatively easy. Production and inventory are prepared according to demand, which is also based on the prediction of the market, license plate development, both domestically and internationally. There are approximately this many.

Especially many people in car manufacturing plants who have never been involved in Robotaxi forget that taxi drivers and ride - hailing drivers do many things other than driving, such as charging, maintenance, cleaning, and handling many customer - related matters like carrying bags. All these things were done by the drivers casually, and for unmanned vehicles, efficient solutions are needed for all these things, and corresponding infrastructure also needs to be established.

This is why ride - hailing platforms claim to have operational capabilities and want to do L4 operations. In fact, this so - called operation is different. Whether it's ride - hailing platforms or taxi platforms, they don't need to do the things I just mentioned because they only need to manage the drivers, and the drivers handle these tasks.

However, these things are non - existent in Robotaxi, so companies doing operations don't necessarily have a natural greater advantage.

For car manufacturers and ride - hailing platforms, Robotaxi is actually a new species and cannot be simply grafted. People say it's a platform - based business, but in fact, what platforms understand is another set of operations. They don't even know what Robotaxi is all about.

36Kr Auto: How does Pony.ai plan to do these things today?

Peng Jun: We need to do all these things ourselves first, establish standards, and build a platform. In fact, there are many technical solutions. For example, how to make the vehicles come back for charging together and how to optimize the peak and valley of electricity consumption.

How can you charge and clean 20 vehicles simultaneously? Plugging in the charger only takes half a minute. How can you improve the efficiency of other tasks? There are many things to do. We are essentially establishing a set of standards. Of course, it's also possible to find a third - party to do these things in the future, but for now, we need to establish our own standards.

36Kr Auto: Currently, with 1,000 vehicles, how many vehicles are matched to one person for ground operations? What is the vehicle - to - personnel ratio?

Peng Jun: Including ground support, remote maintenance, and monitoring personnel, the vehicle - to - personnel ratio is already very low. Labor costs account for a very small proportion of the total operating costs.

36Kr Auto: Is this model based on the current fleet size or a long - term one?

Peng Jun: In the long run, it may increase slightly, but not significantly.

36Kr Auto: How much do you think these invisible operational tasks account for in the barriers of Robotaxi?

Peng Jun: They don't determine whether you can start, but they determine efficiency. If one person can manage 30 vehicles while another can only manage 20, the costs will be different. So, they don't determine whether you can start, but they determine many things.

Technology determines whether you can start. It can be said that 99% of the people are already eliminated. However, after having the technology, these things are also very important because they determine your efficiency.

On the L2 business: Once price competition starts, it's a red - ocean market

36Kr Auto: There is a saying in the industry that mass - producing intelligent driving is too difficult and takes too long for engineering. Is that why Pony.ai has always focused on Robotaxi and hasn't ventured into L2?

Peng Jun: Mass production is actually simpler. The L2 market is definitely closer and more immediate, but it's a red - ocean market. In the end, it's an industry with no profit after intense competition. Robotaxi is definitely a much larger market, but it's also much further away.

From an industry perspective, the technical threshold for L2 is relatively low. Most enterprises have the R & D and production capabilities. The technical paths and implementation methods are similar, and the performance differences between different products are difficult for users to perceive intuitively.

There is also no unified standard. Whether the vehicle needs to be taken over every 10 kilometers or 100 kilometers doesn't make a significant difference in the intuitive experience of ordinary people because in essence, the user (driver) is the backup in L2.

In this situation, automobile manufacturers have the say, and market competition gradually focuses on price. The profit margin of the industry is continuously compressed. We noticed this a long time ago.

36Kr Auto: Around what time was that?

Peng Jun: In 2021 and 2022, when many people started to shift their focus there.

36Kr Auto: The revenue scale of the L2 business is still quite large. I see that Pony.ai's goal for Robotaxi this year is only about $100 million. But if you were to do L2, getting a mass - production project from an automobile manufacturer could reach this scale?

Peng Jun: The market for Robotaxi is getting bigger and bigger and will still be growing in 10 years. The market for L2 may shrink in the next two years. Its shipment volume won't increase much. You can think that the current penetration rate is 60%. Even if it reaches 100%, it's only a double increase. However, the price per vehicle is dropping rapidly, so the market is getting smaller, not bigger.

On automobile manufacturers entering the Robotaxi field: Starting from scratch

36Kr Auto: Taking first - tier cities in China as an example, there are about 100,000 ride - hailing vehicles. How many Robotaxi vehicles would be needed to cover the whole city?

Peng Jun: Our current goal is not to cover the whole city but to reach at least 10% - 20% of the number of ride - hailing vehicles. I think it's feasible and sustainable.

36Kr Auto: At this time, does the manufacturing process or standards of the vehicles need to be updated? Or can they remain the same as the current state?

Peng Jun: We are constantly updating. There are several capabilities, including operational capabilities and production capabilities. The production standards, capabilities, and durability are also constantly improving. Now, with our thousands of vehicles, we are already fully compliant with automotive regulations.

<