Today is actually a media appreciation event, but I think its significance goes far beyond that. I see three very important differences today. I believe this appreciation event is different from previous intelligent assisted driving in China. Before, when we talked about intelligent assisted driving, every company, every product, from my perspective, was at the L2 level. But today, I believe this is also the first in China, a brand new start for L4-oriented intelligent assisted driving. It means that the car everyone drives will be very different from before. This is the first point.
Second, I think a very big difference is that during last year's Spring Festival, we saw China launch DeepSeek. This year's Spring Festival, we saw ByteDance also launch something. This means AI and agents are changing everyone's digital world. But starting from today, I believe China's automotive industry, even China's smart hardware industry, is officially opening up capabilities related to physical AI and agents.
Third, in my first decade of entrepreneurship, I always said that smart electric vehicles initially changed new energy, from gasoline to new energy. We can see that today, when a user wants to buy a car, what do they care about? Range, design, price. In the past decade, software or AI might have only accounted for about five to ten percent of a user's mindset when buying a car. But in this decade, we will see new questions emerge: Can this car drive itself? Can this car proactively serve me? A series of things are starting to happen.
So, I believe that although today is a small step forward for XPeng, I think it is an absolutely huge step for XPeng's new decade. I will quickly share and explain everything in the following presentation. Today, I will report on XPeng's second-generation VLA. Our goal is to cross the first stage from L2 to L4. In this, we believe three capabilities are very important. First, how to make everyone not afraid, but feel at ease, smooth, relaxed, and confident. It's more important than just safety.
Second, full-scenario capability. Although I admit today that our full-scenario coverage might not be perfect everywhere, for example, we are working internally on how to handle a situation where our car meets another car in a narrow alley. How does our car automatically reverse and yield, then proceed after the other car passes? But don't worry, we guarantee to launch this within this year.
Third, how to be highly efficient. This is divided into two points. First, it must drive as fast as a human, so you don't feel it's sluggish. Second, it must be very smart, able to plan over long distances and times. This is what we believe are the three specific experiences for crossing over to L4. In this, we have listed a bunch of various L2 assisted driving scenarios that make you feel uneasy, worried, or scared. We've only listed four very simple ones here, including irregular vehicles, like what to do if a vehicle is overturned, or if a car door suddenly opens on the roadside next to you and someone jumps out, or if there's a very bumpy road section with a pothole, a speed bump, or even some stone bollards. What do you do? Also, in complete darkness at night, if a person dressed in black is walking quietly beside you, what do you do? Please watch as we go through them one by one.
The first one. Let me stand on the right side here. We can see the different performances in handling irregular vehicles. Whether it's a large truck, a vehicle encroaching on one lane or half a lane, or various small cars driving in the wrong direction, single-wheel vehicles, two-wheelers, three-wheelers, or even a mule, it handles them all without a problem. Cool, right?
Let's look at another one discovered by our test staff. When we were driving, someone suddenly pushed a door open. To be honest, I've encountered this once on my way to work. The car was about a meter away from me, and suddenly, a driver jumped out. My car clearly reacted faster than I did because it swerved first, about 30 to 40 centimeters, before I even realized the door had opened and someone jumped out. So it actually sees faster and more accurately than I do. This was an unexpected event that happened during our recent testing.
Everyone can see a car on the opposite side of the road that, due to some situation, crashed into the guardrail. So it stopped in the middle of the road, with debris scattered everywhere. In the video, we not only show this footage, but we also display all the data of what the VLA saw regarding the obstacle and the car. We found that it identified the obstacle from a very long distance away and planned its path in advance based on the accident car's position. Everyone must remember that this path planning is completely real-time, because in this sudden accident, people or other cars might be moving continuously, so it reacts completely in real-time.
Let's look at another one. This filming scenario is even more interesting. It's a section of a bad road. Remember, it doesn't even have a normal paved surface. Sometimes at the junction of urban and rural areas, or in villages, or on village roads, we see this kind of situation: a road with no lane markings, no curbs, nothing. So in this case, we can see the second-generation VLA frequently slowing down and avoiding bumps because it detects the road is uneven or there are objects on it. At this moment, it sees it. It's really great. Honestly, I think it drives better than I do at night, because when I drive at night, I wouldn't be so meticulous in looking at the specific details of the road surface. To be honest, perhaps this year or next year, if manhole covers are stolen from the road, we will be able to automatically identify them in advance.
Okay, this is another very interesting one. This is called a nighttime scenario where you are turning and in your blind spot, there is a person dressed in black walking. What do you do? Would you be startled? I've encountered this myself many times. When turning right, there's a person on the right who was standing still and suddenly moves. When you are turning, you can't see them at all. Everyone can see, look, there are actually two people. But in that moment of turning, in the dark night, in the rain, with them wearing black clothes, you completely don't see them. At this moment, you will see that our car's right-side camera has already spotted them in advance. This video makes me firmly believe that the ultimate safety of intelligent driving will definitely far surpass that of humans.
Let's try to watch the next video. I see many friends filming. Okay, let's see what this is. When a small dog appears at night, the car slows down. Awesome, right? This video left a particularly deep impression on me. In 2017, when I started to think about whether autonomous driving was a necessary endgame for the next ten or twenty years of cars, I went to Silicon Valley. I remember I took several autonomous driving tests at Silicon Valley R&D companies in 2017. Some drove quite well, to be honest, but they ran over a squirrel. I was particularly impressed. First. Second, they were actually only able to perform relatively well in simpler daytime environments. So at that time, I remember coming back and telling our internal development colleagues, I said, in the future, can we identify various animals, cats, dogs, and if we go further, horses and all kinds of animals? Do you know how software used to identify what kind of animal it is? There are countless combinations of animals in the world. You might have a group of cats and a group of dogs combined, so you simply cannot judge whether it is a cat or a dog. Only with a completely new AI plus vision, I believe, can we do it well.
So in this regard, how to evaluate a good intelligent driving system, whether it is safe, reassuring, smooth, and confident, and you won't be afraid, I think we have considered four points. First, does it drive abruptly? For example, sudden braking or sudden acceleration. Second, does it sway? That is to say, when it changes lanes or turns, will it make me sway left and right? Third, is it bumpy? That is to say, if there is a speed bump ahead, thump, a jolt, or a pothole ahead, boom, a jolt. This is the X, Y, and Z axes. Fourth, is the entire journey smooth? That is to say, the overall planning of braking and starting, the pre-buffer amount. So for this, we made a very interesting piece of software called the XPeng Intelligent Driving Comfort Evaluation Software. 'Spilled it?' This 'Spilled it?' Let me tell you a background story first. You know, in the past, drivers for leaders had to take an exam. What was the exam? They had to place a cup of water, later it became a cup of coffee, full. You drive from point A to point B and see how much spilled. If it's completely not spilled and the efficiency is high, it means you're okay. If it's completely not spilled but you drive like a turtle, fail. So from another perspective, an extremely good driver is highly efficient, very smooth, very reassuring, and doesn't spill coffee or water. So we specifically made a software for this. Welcome everyone to scan the code and try it. We just want to let different cars, whether you drive it yourself or other intelligent driving systems, or XPeng's VLA, be able to compare. This comparison is for us to let our own VLA, we expect to reach the level of a human expert driver, reach the level of the most professional human expert driver, reach the level of the most professional local expert driver. Because many expert drivers from Guangzhou have also gone to Beijing, or from Beijing to my hometown in Huangshi, Hubei. You will definitely feel you are not used to driving because you are in a completely new environment, or if you go from China to the United States or Germany, you will also find you are not used to it. It is a local expert driver familiar with the road conditions who is the real deal. So the intelligent driving industry can finally have a benchmark score.
Let's look at some data. This is our second-generation VLA compared to a previous L2 intelligent assisted driving system. They drove the same amount of time on the same road section for a set of evaluations. In this evaluation, for the original L2 intelligent assisted driving, what was the most painful thing? Too many sudden brakes. That is to say, several sudden brakes made people feel scared, and because of the fear, they didn't dare to use intelligent driving anymore. But on the same road section, our second-generation VLA drove very smoothly. So I am particularly happy. I believe XPeng's second-generation VLA's first goal is to solve the problem of making a national intelligent driving system that even moms can drive with peace of mind.
I heard the best feedback from a mom this time. She said that through trying it, she said intelligent driving finally drives better than herself, better than her husband. She said, 'I dare to buy a car now.' She said, 'Before, when my husband and kids weren't around, I didn't dare to drive or park.' She said, 'Now with XPeng's VLA, I drive out myself. It drives better than me. I am very reassured. I dare to go out and play. I dare to take my friends out for tea.' I am very, very happy. This is also XPeng's goal. We must achieve this.
So we just talked about making it so that even moms can drive with peace of mind. This is the first point of the second-generation VLA. The second point, let's look at full-scenario capability. It can drive anywhere. This is very important. In the past, many L4 intelligent assisted driving companies said their MPI or safety takeover rates were very high. I never really believed it because they drove very slowly and didn't go to many places. They only went to main roads and highways. They didn't drive on any other roads. We often take various L4 cars, and we say from point A to point B, it takes a long detour because it avoids all complex roads. But XPeng's VLA, whether compared to L4's total global test or the current L2, our goal is to be able to drive on all roads.
In this regard, I want to explain to everyone. You can see that many L2 assisted driving systems today, when they are sold, say they support highways, or even city memory navigation. What is memory navigation? Memory navigation is just the city's main roads or expressways, which are simple and similar to highways. Essentially, it can only drive on highways. It also knows that in many sections of the city, it might not be able to drive, or it's unsafe to drive. But today we can see, from the city's main road to the side road, if we stop on the side of the side road, can we start in P-gear and activate autonomous driving without having to drive out ourselves? If we drive to a company's or family's underground parking lot, can it drive in automatically? If on very narrow road sections, for example, hutongs in Beijing, or old urban areas in Guangzhou, does it dare to drive automatically? If there are many road construction closures in China, or temporary objects placed, can it drive through? If all these can be driven, is it possible to drive on rural paths, township roads, or even snowy roads? I think this is the full scenario. Today I also want to say that XPeng's VLA cannot drive in all scenarios, or not drive well in all. For example, I think our highways, our urban areas, let's just say urban areas. In urban areas, if we drive today, I think other assisted driving systems are probably at 70 points, and we might be at 99 points. But on small roads, they might be at 10 points, and we are at 75 points. In parking lots, they might be at 0 points, and we are at 60 points. But I am very confident that within this year, all our roads will be driven to 99 points, just like the main urban roads. Let's look at a series of short videos. This is the first one called P-gear activation and start. We can see that in this generation of VLA, we have developed the ability to activate NGP directly from P-gear, whether parked on the roadside or in a parking lot, and start driving. But I also want to explain to everyone that if the cars in front and behind are very close, this version does not support it yet. But please look forward to it, we will solve it very soon in a later version this year. This is P-gear start. My goal is for you to walk from your home to the parking lot, get in the car, and it drives itself to your company's parking lot or your school's parking lot. This is true autonomous driving. Driving many round trips in between without a single takeover. This is safe and excellent driverless driving, even when encountering various abnormal situations.
Second, let's look at this. This is a passage through a small road in an urban village in Guangzhou. You can see that the distance between these two cars might be less than 10 centimeters. Because from the rearview mirror angle, in the past, you would never dare to let go. This would definitely require a takeover. But this one, remember, this is played at original speed, without any acceleration, and it drove straight through. Actually, in our daily lives, similar scenarios are very, very common.
Okay, let's look at a more interesting one: a rural village road, and it's a nighttime village road. Look, why is it signaling left? Because it detected a pothole on the right, so it turned left on its own. In this situation, we can see that with the second-generation VLA, it can completely navigate rural paths with no lane markings on the ground, obstacles on the ground, and potholes, just like a human driver, accurately identifying and avoiding them.
Okay, let's try to watch a more challenging one. This is a nighttime scene. See? Many cones are placed to block the road. Look at how this car is driving. Look. Okay, at this moment it drove to the right, right? Because it saw the cones ahead were placed relatively straight, and then it discovered another blockage ahead, so it drove left again, bypassing the cones ahead, and passed through completely smoothly. To be honest, I think such a scenario today is already a full process of navigating small, nighttime, irregular objects, extremely smoothly. This is completely without acceleration.
Let's look at a scenario that makes all of us feel scared: driving inside a night market at night. You have to know what's in a night market: many irregular objects because many stalls are set up, second, lots of people, many bicycles, all kinds of miscellaneous items, children running around everywhere, and the lights are very chaotic. At this time, how to drive in a nighttime scene like this? We can see the second-generation VLA is also very smooth.
Okay, let's move from nighttime to a heavy rain scenario. This is what we encountered recently during testing: a downpour. In the past, many intelligent assisted driving systems, when it rained heavily, we could see they would require a human takeover. Actually, it's because they couldn't judge the entire environment. When writing rules with software, they found many problems arose. So if it actually rains heavily, people will say it affects vision, but it affects LiDAR even more. So in a visual scenario, even in heavy rain, the second-generation VLA can still be made extremely smooth and reassuring.
Okay, let's look at a more interesting scenario. We often encounter situations where we don't know where we are going. For example, we leave a parking lot to go home, we start driving and then navigate. We leave a restaurant after eating, we usually start driving and then navigate, right? At this time, you haven't set a navigation destination yet. It can drive autonomously without setting a destination. Awesome, right? During the Spring Festival, I drove this feature in Guangzhou. I drove for over an hour, wondering where it was going because I hadn't set a destination and didn't know where I was going. It just drove around Guangzhou on its own. It was like a leisure mode. I myself felt this situation is really great. When you leave a community or a parking lot, you first expect it to drive out automatically, and then you fumble to set a navigation destination. So at this time, it can be very safe and reassuring. And this roaming in parking lots and roaming in urban areas will be released in this update. Roaming in underground parking lots, I expect it to be in this update as well, but at the latest, it will be in the next update. We will definitely usher in an era of full-scenario roaming. No matter if you set a navigation destination or not, it can drive.
Let's look at another one. If the parking is a mess, even humans are confused. Intelligent assisted driving activated. Pretty good. Driving slowly on the narrow road. Clear ahead. Driving quite fast. You're quite confident. Why is this car parked like a maze? It's too narrow here. Steady. Hey, yes, yes, just like that. Very cautious. Are you afraid of scratching my car? Okay, okay. An experienced driver. Your steering wheel is also quite smooth. Impressive, right? I'm telling you, even if every one of us, I believe many here are professional experienced drivers, but I want to say, if we were to drive out of such a complex, randomly parked parking lot, we would also drive quite fast, looking left and right, driving slowly. But I think it drove, maybe a little smarter than my brain. On a very narrow dead-end road full of randomly parked cars, it found the exit on its own. So it's really different. What does it mean to be able to drive in all scenarios? Besides our standard highways, urban main roads, urban expressways, all park roads, hutongs, parking lots, construction zones, blocked roads, night markets, dead-end roads, unmarked rural paths, etc., urban village gates, side roads, heavy rain, nighttime, it can drive in all of them. This is our full-scenario capability.
Now, let me share the third point. XPeng's second-generation VLA makes experienced drivers, professional drivers, enjoy driving it. You know, experienced drivers, I am one of them, I always feel you drive too sluggishly. I want to arrive faster. I feel my planning ability is stronger than yours. I feel I can squeeze into a gap, and it's very safe. Don't worry, there will absolutely be no collision. So experienced drivers have higher requirements for a car's autonomous driving than we do. First, efficiency. Second, being very smart. In the past, many L2 assisted driving systems, in order to reduce takeovers, would first slow down in difficult areas, and second, not drive at all. It would severely sacrifice traffic efficiency. At this time, I often get anxious. I often say, why don't you turn left? There's no one on the left. Why are you stuck here? So encountering difficulty means slowing down, and if you can avoid changing lanes, you don't. The prediction ability is extremely weak. This was the case in the past. Our second-generation VLA, in terms of traffic efficiency, I want to say, has initially reached the level of an experienced driver. I say 'initially' because I believe we will make huge improvements this year. We can see this is a recent test on a real road section. We took the second-generation VLA with navigation, L2 assisted driving, and L4 robotaxi. You might be surprised, why is the L4 robotaxi's route a bit different from yours? Because we told it to go from point A to point B, it can only specify one road, so the road it drives is not exactly the same as ours. It took an easier road, while we took a road with a relatively balanced complexity. So we can see the navigation estimated time is 44 minutes. The second-generation VLA is close to it, while L2 assisted driving and L4 robotaxi both take several minutes more. So the second-generation VLA's 'chauffeur', a friend told me a story. In January this year, I once invited a friend who has driven for many years. At that time, he had no idea we were internally testing the VLA. I asked him to activate it and drive a short segment. He drove for over an hour, and finally he said, 'Mr. He, I think how many more years can our drivers drive?' Why? Because he said, first, he felt his greatest driving skill was his right foot, meaning braking, turning, and accelerating are all very smooth. He said, 'I feel my right foot is not as smooth as your intelligent assisted driving.' Second, he said, 'I feel I know Guangzhou very well, I'm familiar with Guangzhou, but from another perspective, it seems to be about the same as me.' So from another perspective, comprehensive driving efficiency is extremely important. We currently believe we have improved comprehensive driving efficiency by 23% compared to current L2 systems, but I believe there are even higher expectations. I believe in a few years, XPeng's intelligent assisted driving will not only drive well but also drive extremely, extremely smartly. Let's use two short clips to see what extremely, extremely smart looks like.
We can see this is a very interesting short video from our testing. Our car drove to a section and suddenly stopped. Why? Wow, it turned out an ambulance was approaching from behind. It just stopped on the side of the road. After the ambulance passed, it started driving again. This was completely randomly encountered during road testing. To be honest, it's really hard to capture.
Okay, let's look at another very smart point. During our testing, we encountered police checking for drunk driving at night. Do you know what XPeng's VLA did? It drove over on its own, very honestly stopped, and blew into the breathalyzer. The police officer said, 'Go, go, go,' and it drove off. Smart, right? To be honest, we also encountered some not-so-smart situations. Our team told me, they said, 'Mr. He, when we first started working on VLA, a very painful point was that the car would often automatically pull over and stop for no reason.' They went to investigate why. Because it happened that someone nearby was hailing a taxi, and our car understood and stopped on its own. You shouldn't stop in that situation, right? So we need to identify when you should stop and when you shouldn't. But I also want to say, this is the brand new capability that AI brings us, and it's our brand new VLA.
So today we introduced the national intelligent driving system that even moms love to drive. We believe its safety, smoothness, and comprehensive capabilities, the results we have tested so far, we believe are 5 times or even more improved compared to all the top L2 systems in the industry. We believe we will achieve full-scenario coverage from start to finish. This year, we will make basically all coverage as good as the main roads. Today, our comprehensive driving efficiency has improved by 23%, and this is just the beginning. We believe this will definitely open up a brand new moment. I want to say that recently, you can see that in the past decade, all new car-making forces started out very painfully. Gasoline car companies had huge monthly sales, but new energy ten years ago only accounted for 4/1000 of today's total industry sales. But I want to say that in the new decade, we will definitely see a complete change in the evaluation criteria for cars. At that time, whether in China or in the United States, you rent an XPeng car or buy an XPeng car, you can tell it, 'I want to go to Canton Tower, I want to go here or there,' no matter where you are, you can use Chinese, and you can completely achieve fully autonomous driving, extremely safe, taking you to your destination. I believe this is an extremely important moment. This is a brand new moment for physical AI. I believe agents will definitely become the important moment for cars in the new, not to say ten years, but the new three years. In the past, many companies said you should buy software from Tier 1 suppliers. I want to say that today, cars are undergoing full-domain integration. Today, when a car's core competitiveness increases with agents and AI, only full-domain independent research and development can make this cross-domain integration possible. So I believe today's moment is a very, very important one, because we can see that XPeng's own VLA, its acceleration is really too fast. I once sighed, I said the progress we used to make with software-written autonomous driving every year, now it doesn't even take a month, even two weeks to achieve. So, please believe, March, April, May, June, July, August. August is the big test for me and our autonomous driving team, Mr. Xie Ming. In August, we say we want to achieve in China the effect that Tesla has in Silicon Valley. Remember, achieving the effect in China is far more difficult. I believe we can definitely achieve this.
I believe that fully autonomous driving, after policies and regulations open up, will definitely allow everyone to get home safely after drinking at night. I believe it will also allow our parents and our children to go out in the future without us having to drive for them, letting the car proactively serve them. I believe this era will come very soon. So during this time, many friends haven't even touched our car, have only seen videos, and only heard our introduction. I want to say, starting today, we are inviting 200 professional media outlets from around the world for test drives. Starting today, we welcome comparisons. We welcome you to use the most challenging road sections to test us.
Additionally, we also plan to do a west-to-east intelligent driving challenge across China in the future. We know that during the past Spring Festival, many friends drove hundreds of kilometers round trip, and some might have driven thousands of kilometers to their hometowns. You have to know, that kind of driving is truly a painful ordeal. Every four hours you drive, regulations suggest you take a break. It's very, very tiring, and you have to pay attention to safety. And it's about a family's happiness. Our goal is to make your next family trip safer, more reassuring, and more worry-free.
Users don't need to worry, friends please rest assured. Starting from March 11th, our VLA 2.0 will be open for experience at all our stores nationwide. I hope that whether you have driven intelligent assisted driving before or have never driven it and want to challenge, learn, and experience it, you are welcome to come to our stores. At the same time, remember to use our 'Spilled it?' software to see if it drives better than your son, or your husband, or your dad. Welcome to challenge them.
I am also very happy to announce here that our robotaxi, which uses the same software as our second-generation VLA, our test vehicle is our GX, has already completed the entire operational process of starting from a standstill, driving autonomously, pulling over to stop, picking up passengers, starting the meter, and all other operations, with no one in the driver's seat and no maps in the park. I hope to open up more test drives for more friends in Guangzhou as soon as possible this year.
There is also one more thing. Because XPeng's VLA, from its inception, has considered how to globalize. We are not willing to write rules just for China. We know many L2 assisted driving vendors have written countless rules specifically for Beijing, Shanghai, Guangzhou, and Shenzhen in China, because media test drives are basically in these four cities. So their software, when taken globally, has to guess rules and rewrite rules. But sorry, they don't know the local regulations. Sorry, they don't know how local drivers change lanes, turn, park, or overtake. They are in great pain because they use a lot of local high-definition maps, especially at complex scenarios like intersections. These cause current L2 systems to look good but have a very low ceiling. So I am very happy that we took this VLA to other countries' cities for some testing. Everyone can see this map. We completely used no local data, did no training, and it can still drive, and drive very well. I am really very happy. I look forward to seeing it overseas very soon, in the near future. When will we see it?
I expect it to be in 2027. In 2027, we expect to officially begin large-scale global delivery of XPeng's VLA. And it will be the full version, not a stripped-down version. As I just mentioned, the UN's DECAR traffic regulations have recently announced that in 2026, drivers will be allowed to take their hands off the wheel during assisted driving. In early 2027, it is very likely that high-level assisted driving will be opened up. So I believe the opening of policies and regulations has already begun, and our technology will definitely arrive. So, let's put some pressure on ourselves. I expect the era of fully autonomous driving to arrive at XPeng within the next one to three years. I expect everyone to feel that at that time, it drives more safely, reliably, smoothly, and efficiently than you do, and you have confidence. Please stay tuned.
Besides XPeng using our second-generation VLA ourselves, we also very much welcome more OEMs to use our second-generation VLA. I'm also happy to share with everyone that Volkswagen has become our first launch customer. In fact, there are other customers currently in discussions. I believe the best intelligent assisted driving will be a standard feature of intelligent car agents in the future. Cars will transform from passive driving to fully proactive service agents. This is definitely a milestone. At the same time, I'd like to share some great news: the first step in our embodied intelligence deployment is in our VLA, also on our own cars. But soon, robotaxis will meet everyone. After that, our robots... I believe we will be the first company in China, and even globally, to not only research VLA but also mass-produce and deploy it. We will definitely show everyone the effect of our VLA in full-scenario embodied intelligence. It's truly beyond imagination. In the past, I wouldn't have dared to imagine that a humanoid robot could generalize freely, communicate, move, and operate. But I believe that in the not-too-distant future, it will definitely be possible. So many friends might ask, 'XPeng, how can you achieve this?' So today, I want to first show a short video, and after the video, we will invite the head of our second-generation VLA, Liu Xianming, to introduce to everyone how our VLA was made. Let's play a short video first, showing a few of our car owners taking our VLA home for the New Year, filming real scenes. It's very cool.
Please welcome Liu Xianming, head of XPeng Motors' General Intelligence Center.
