I used to be strolling by Corridor B1 of the Munich Exhibition Centre in Germany, the place the 2025 IAA Mobility passed off. Lots has modified up to now years for one of many world’s oldest motor exhibits. Initially, it’s not known as Motor Present anymore, as in 2021 it rebranded from Internationale Automobil-Ausstellung to easily IAA Mobility and moved from Frankfurt to a much less shiny location in Munich, reflecting the change the automotive business is present process.
I’m passing a large Volkswagen sales space, by far the largest within the corridor and possibly on the entire present. I catch a glimpse by the gate main contained in the sales space: a bar, tables filled with meals, and crowds of sellers, KOLs and journalists feasting on the generosity of Germany’s largest carmaker. It’s clear who has a house floor.
However my vacation spot lies elsewhere. Proper subsequent to Volkswagen’s fort, a smaller sales space has attracted a big crowd. Individuals collect round a strikingly inexperienced, futuristic-looking sedan. It’s up to date P7, which had its international debut right here just some hours in the past as The New P7.
The proximity of the Xpeng-Volkswagen sales space shouldn’t be a coincidence. In 2023, Volkswagen sought Chinese language companions to speed up its EV improvement cycle and located them in Xpeng, to which it invested over 700 million USD for about 5% share. Xpeng will present its E/E platform, and each automakers will collectively develop two Volkswagen-branded mid-size pure electrical automobiles, which is able to launch in 2026.
“Meet us in entrance of the SEPA platform,” posted the Xpeng consultant within the group chat as we have been finalising particulars of the interview with Tingting Yuan, often known as Candice Yuan, director of the Autonomous Driving product at Xpeng.
For Xpeng, its autonomous driving is a key differentiator within the fierce China EV competitors. Nio has it’s premiumness, BYD price-performance ratio, Li Auto is the chief in vary extenders, Zeekr and Lynk&Co attempt to appeal to younger and hippish clients, and Xpeng has it’s ADAS (superior driver-assistance system) and autonomous driving software program. Final 12 months, the corporate additionally began closely selling its AI capabilities. “Xpeng is constructed with AI in its DNA,” stated He Xiaopeng, firm CEO, throughout a media briefing in Munich.
In an interview with Candice, we lifted the lid on a lot of these subjects.
Earlier than you joined Xpeng, you labored at Alibaba, the place you have been liable for enterprise operations of unmanned logistic automobiles, Little Donkey and Large Donkey (editor’s word: low-speed self-driving automobiles and unmanned heavy-duty vans). What expertise did you carry to Xpeng from that?
It’s an enormous distinction from Alibaba Damo Academy. In Alibaba, we did logistics, which meant no driver or passenger within the automobile. So we didn’t have to fret about consolation. We may solely give attention to the protection half.
However for the human drive L2+, it is vitally advanced. Firstly, drivers need effectivity. For instance, if ADAS do the correct factor, however the automobile is slower than the driving force expects, he’ll take over. One other instance – the system would possibly calculate that the scenario is secure sufficient to go. However the driver thinks it’s too slim or too shut, and he’ll take over. The third instance may be that if the automobile accelerates too quick or too gradual, the driving force is not going to really feel comfy and, once more, take over.
Consequently, a higher stage 2 autonomous driving for passenger automobiles is extra advanced in some areas than the L4 we developed in Alibaba for unnamed automobiles. The driving force cares about feelings and emotions, so we have to assume as he thinks. We have to focus extra on expertise.
So there’s a big distinction between creating L4 for unnamed business automobiles, which is definitely simpler in some methods than L2+ for passenger automobiles with a driver?
I can provide yet another instance. Possibly you might have already taken a journey in Waymo or one other robotaxi. If a robotaxi goes the fallacious method, it’ll proceed in that route, then flip round and observe the brand new navigation. However in stage 2, for those who go the fallacious method, the driving force will take over. So .. it’s forbidden to go the fallacious method.
What would you want to realize in Xpeng and its autonomous driving system?
I believe it’s the identical for each workforce within the autonomous driving business – to realize stage 4 and stage 5.
I believed you’d say this.
Sure, everybody. However I imagine Xpeng has an important benefit in doing that. I’ve been within the autonomous driving business for greater than 10 years. On the earlier time, I might assume, who will take extra benefit of it, the automaker or the ADAS maker? Identical to the Android system and the smartphone makers, or Microsoft Home windows, and, for instance, Dell laptops. {Hardware} vs software program.
However these days it’s clear that {hardware} and software program can’t be aside. They should be linked collectively. And the nearer the connection will get, the sooner it goes.
So Xpeng is the automaker. Now we have a workforce of greater than 1000 workers simply targeted on autonomous driving. And we are able to do very nicely in HW and SW. We additionally create our personal AI chips ..
…I’ve questions for that later. Let’s discuss lidar a bit now. Three months in the past, you posted an evaluation of its professionals and cons. Xpeng was utilizing lidar for seven years, and final 12 months, you eliminated it from the P7 sedan and all of your future fashions. Was it a tough determination, and the way lengthy did it take?
Truly, from the times that we used lidar, we have been constantly fascinated about eradicating it for a few years. We knew we must always take away it at some point. However up to now years, we had no capability to do this.
Do you imply software-wise?
Sure.
Obtained it.
These days, we’re extra assured that eradicating lidar was a good selection. Why? Our new AI system is predicated on a big language mannequin primarily based on many information. The info are largely quick movies, minimize from the street whereas the client is driving.
It’s a quick video, like 10 or 30 seconds quick. These movies are enter for the AI system to coach on, and that’s how XNGP is upgraded. It’s studying like this, it’s studying from each automotive on the street.
The lidar information can’t contribute to the AI system.
Why?
As a result of there’s solely visible enter, we name it VLA – imaginative and prescient, language, motion. Lidar information are totally different and may’t be absorbed by the AI system. That’s the reason our system grows very, very quick, as a result of we are able to practice it on a lot street information.
So, when there’s an accident or some drawback on the street, the AI mannequin will study from it and replace itself?
Not solely from unhealthy circumstances, but in addition from good circumstances. We name it good driver behaviour: You possibly can flip lef,t you’ll be able to flip proper, there are all the time 1000’s paths you’ll be able to take. However which is the easiest way? We are going to see how good drivers deal with it and practice the system from them.
I wish to study extra about XNGP (Tesla FSD different, editor’s remark). I did analysis about it, however I’m nonetheless confused about some elements. Each Xpeng automotive in China is now geared up with XNGP?
Sure
Is it unbiased totally from HD maps, or is there some hybrid resolution?
There aren’t any HD maps. Xpeng is the primary automaker to take away HD maps from it’s NOA system
Final 12 months, He Xiaopeng stated that in 2025, you’ll start the rollout of XNGP globally. How does it go?
We’re the primary to have approval for distant autonomous parking in Hong Kong. This is step one for international rollout.
In Europe, we have already got ACC and LCC (adaptive cruise management and lane centring management, a system that helps maintain the automobile within the centre of its lane, editor’s word) for all automobiles, so we take away some burden from clients, and so they can calm down. It’s a fundamental ADAS. We don’t have XNGP in Europe, and it’s linked to coverage.
We observe the ideas of EU insurance policies just like the DCAS (Driver Management Help Methods) regulation.
If the coverage is prepared for us to carry our system right here, our giant mannequin will be shortly utilized wherever on the planet.
So the coverage is the largest hurdle to bringing XNGP to Europe?
I believe it’s a very powerful cause.
Now the query everyone might be asking you – when will you carry XNGP to Europe?
I can’t say, it relies on coverage progress. However we’re very desperate to carry XNGP to the EU.
How many individuals are working at Xpeng’s autonomous driving workforce?
Greater than 1000. Now we have hubs in China, the US and Europe. The key workforce is in China, and abroad is a number of hundred.
What number of years do you assume till L5 is achieved?
L4 is similar to L5. L4 signifies that in most areas you’ll be able to drive it robotically. If we obtain L4, L5 could be very shut. Crucial step is L2 to L4.
To succeed in L4 relies on many circumstances. First, it’s coverage. The federal government coverage influences it very a lot. In China, L4 and L5 are solely permitted in some small restricted areas.
China is the primary market which opened to L2+. L2+ means the XNGP can drive wherever when the driving force is able to take over each minute. So for L2+ coverage, I believe China is essentially the most forward. So if we are able to see any nation with an open angle to L4 and L5, the progress shall be sooner.
The second factor is software program capability. Xpeng could be very near reaching L4. Simply right now (Sept 8, editor’s word), He Xiaopeng introduced we’ll begin mass manufacturing of L4 robotaxi automobiles subsequent 12 months.
Do you concentrate on utilizing an imaging radar like Mobileye?
For now, we rely primarily on the digicam. However we don’t set limitations. Nevertheless, as of now, there isn’t any dialogue about including this sensor.
We observe the primary precept – folks drive, utilizing eyes, arms, and toes. They don’t seem to be Superman who makes use of lasers. So we’ll observe this.
You developed your individual ADAS chip known as Turing AI. What was making it?
When creating XNGP, we discovered that we needed to join the {hardware} and chip and the software program extra intently. We wanted to supply particular necessities for the chip, however third-party suppliers have been unable to revamp it to fulfill our specs. Our in-house developed chip shall be extra usable for our system, which is able to assist us go sooner.
I might conclude with questions, what’s your day by day driver, however out of your earlier interviews, I do know it’s G9.
Sure, as a result of I like huge SUVs. Possibly I’ll change it subsequent 12 months for a a lot larger one.
Wait..Xpeng has no larger SUV than the G9.
Nicely, He Xiaopeng talked right now in regards to the L4 mass-produced robotic automotive, which is able to launch in 2026. Possibly I’ll purchase this one. But it surely’s nonetheless a secret.
I see. Thanks for the teaser. And your time
Thanks too.
