Name: Pony.ai CEO on the future of autonomous driving and AI
Uploaded: 2025-07-20T18:35:32.000Z
Duration: 18 min 26 s
Description: Thousands of YouTube videos with English-Chinese subtitles! Now you can learn to understand native speakers, expand your vocabulary, and improve your pronunciation...

Let's talk about some of the challenges to drive this car.

Reflexive pronouns

Modal verbs - B1

With any networking glitch, you may run into a safety issue.

Founded in 2016, Pony AI has become one of the world's leading autonomous driving companies.

It has operations across the US and China and recently carried out an IPO on the NASDAQ.

Pony AI creates the technology behind driverless cars and seeks to partner with automakers.

While real-world trials of driverless cars are underway, mass adoption hasn't quite happened yet with challenges including public perception and regulation.

Let's get into it with James Peng, the CEO of Pony AI.

James, thanks for joining me on this special episode of Beyond the Valley, recording in front of a live audience here at Converge Live in Singapore.

It's great to have the opportunity to actually talk to you again.

Yeah, it's been a long time since I was living and working in Guangzhou.

Just talk to me about the driverless car journey so far since your founding in 2016 to where you are now.

Sure, probably I'll paint the picture of what's now.

Actually, in several of the cities for us, on the open streets of Beijing, Guangzhou, Shenzhen, and Shanghai, the local residents can already use our app just called Pony AI, or they can use the third-party ride-hailing apps to actually hail in a ride, just like you are using an Uber or Grab.

And the local residents, while you hail in the vehicle, the vehicle will come in without the driver behind the wheel.

And it can take you wherever you specify in your app.

So I think the experience itself is already quite just like a typical taxi.

On those streets, typically we have 20 to 25 paid customers, paid rides per day, per vehicle.

So we can see it's already a popular service in the local area.

I think now the challenge is how we can quickly scale them up.

So I think about three years ago, there was still a driver behind the wheel in the robo taxis in China in particular.

But now, like you said, you can hail an autonomous car via the app.

How did you go so quickly from having a safety driver through to completely driverless?

And specifically, I guess in China, you've got operations in Guangzhou, Beijing, Shanghai, and Shenzhen.

What has happened in China that's allowed this rapid development?

Yeah, it seems rapid from the get-go, from a thousand miles away, but actually it has took us six years from the first trial to become fully driverless.

Our fully driverless cars first rolled out at the end of 2022.

So it's about a bit more than two years now.

Of course, behind the scene, it has been countless miles of testing, the improvements of AI algorithm, and of course, most importantly, a deep collaboration with OEM so we can fit the vehicles with sensors and AI algorithms and make them safe on the road.

Our record actually show that the average safety of our autonomous driving vehicles is about a magnitude, almost 10x safer than a typical human drivers in the same place.

What do you think are the main differences between China's approach at this point to autonomous driving and robo-taxis versus what's happening in the US?

They're the same vehicles with sensors and AI algorithms.

But I think there are two major differences.

One is the driving condition is quite different.

I think in most of the major cities in China, the streets are typically more crowded and there's more pedestrians than the cyclists.

So I think the predictability of the road conditions is a bit harder on the roads in China.

Another major difference, I think it's the supply chain of all the sensors and the chips and the vehicles itself, especially the battery-powered electrical vehicles.

I think because of the ecosystem is maturing, the cost of autonomous driving vehicles in China is generally much cheaper than what's in the US.

So I think those are the two major differences.

James, the world's been talking about AI for the last two years nonstop.

What we're seeing with your technology is the application of AI in action, right?

Just run us through, you're not making cars, to be clear.

You're providing the technology that allows these cars to be driverless.

What is it, the technology you're providing, and also how is AI enabling the car effectively to drive itself?

We are equipping the vehicles with a virtual driver.

And then we need to predict their behavior.

And then we also need to figure out our actions.

So in short, the driving itself is an iteration of perceiving, prediction, and then making the decision planning.

What's behind all those is we modify the vehicle with sensors.

We equipped with computing power and also AI algorithm.

That's like our brain to make the decision.

So with all those, and then we let the vehicles on the road continue to practice.

So it's like ourselves learning how to drive.

And with all this, it's gradually improving.

And one of the key developments over the past two years has been the rise of generative AI, foundational AI models, large language models.

Do any of these new advances in models have an impact on your tech stack and the way your cars and your technology can approach autonomous driving?

Actually, a lot of the similar algorithms and the similar approach, we actually have adopted them even before the general AI becomes a buzzword.

For example, some of the algorithms we use, the end-to-end training, where we use a data-driven approach to feed the vehicles to learn how to drive.

Another example is we use reinforcement learning to help the vehicles.

So the general or the basic idea is quite simple because driving self is not a rule.

The hardest part of driving is how the vehicle interact with the other cars and the pedestrians and the cyclists.

So the driving self is really an interaction.

That's really hard, but the interaction self actually can be trained with reinforcement learning.

The vehicle self is playing a game, how we can ensure the safety of ourselves and everything around us in a mutual understandable way by signaling our intention with other vehicles.

So that iterative process and that approach and that thinking is in general shared with all the other AI applications.

Yeah, and just on the sort of algorithm and model side, one of the key developments has been the improving efficiency of those models.

When you have that conversation, of course, deep seek always comes up in that, but it's about efficiency.

You know, a key part of your technology stack is the way that the AI on board is also interacting with the cloud and the network.

But there is also a push for more and more AI applications to become on device, processed on board for more efficiency, greater safety in terms of data, et cetera.

How are the more efficient models getting you to that stage?

Actually for the autonomous driving, it has always been focusing on the edge computing because I think that it's a real time application that requires almost 100% safety requirements.

That means anything relying on communication with the server means that with any networking glitch, you may run into a safety issue.

So it has always been on board computing, on board algorithm needs to be the one that ensures the final safety.

So the optimization and improving efficiency has always been our focus for the last eight years.

So the key challenge is really how do we use the limited computing power within the vehicle and then have the AI algorithms to really compute as efficient as possible?

Because it's not just about computing power, it's also about if we use less electricity, less power on our AI algorithm, that means we can give more range into the vehicle because we share the same battery pack with the vehicle.

James, let's talk about some of the challenges to driverless car adoption as it stands.

Subtitles ListPlay Video

Pony.ai CEO on the future of autonomous driving and AI

episode

perception

approach

critical