https://insideevs.com/news/779701/xpeng-tesla-china-ai-robots/
Xpeng Is Running Tesla’s Playbook For World Domination. It May Just Work
Robots. Self-made chips. Self-driving cars. Xpeng’s plan sounds familiar, with just as much to prove.
Xpeng remains unclear about the commercial viability or path forward for the robot itself, aside from a vague promise to mass-produce the androids and sell them in 2026.
*Xpeng’s robot project may be unclear, but like Tesla, it clearly wants to be more than just a car company. *
The brand calls itself a “tech startup,” although at 11 years old and more than 1 million vehicles produced around the globe, I think we’re a little well beyond the title of “startup.”
Xpeng has moved forward to occur a market space Tesla has eschewed,
Xpeng’s roughly $18,000 Mona M03 sedan showed up. In a little over a year, Xpeng has sold 200,000 examples of that car, and has plans for the model to enter other markets outside of China.
Recently, it unveiled a “Max” version of the Mona M03, which adds Xpeng’s Level 2 assisted driving features, at the time not super common in such an affordable car.
Despite those wins, it’s hard to stand out in the very crowded Chinese automotive space. That’s why Xpeng is clearly looking beyond cars.
Xpeng’s World AI Day started off with an announcement, a “new invention,” it purported, called “Physical AI.”
Xpeng sees physical AI as the manifestation of its large data models, making sense of the world. The founder and CEO of Xpeng, He Xiaopeng, wants the self-driving tech to have similar reasoning and response to that of a human, and less like a machine.
To Xpeng, human logic is effectively translating “visual information” to action, whereas a typical AI model will take visual objects alongside language to make sense of the world. In layman’s terms, AI models are constantly taking in data, whether that’s actual text, or information from via camera or LIDAR. It’s then organized and interpreted by the model itself. Keep in mind that this all happens while the car is driving, so new information is constantly being absorbed and processed.
Now, add in the fact that the car (or robot) has to navigate a world that is changing in real time; it’s a unique challenge for an AI model, especially compared to other AI models that don’t have to deal with changing information in real time.
Your brain does this all of the time, so it deserves more credit than you think. Getting a machine to do the same is very hard. To get something that’s usable, Xpeng says they essentially had to pare back how much data and what kind of data is interpreted by the AI model itself.
Xpeng calls it VLA 2.0, short for Vision, Language, and Action—although with the 2.0 version, the Vision and Language aspects have been combined.
Essentially, in earlier VLA models, Xpeng would process both vision (objects), along with everything else in the background—most notably, language and text. This overabundance of data is why Xpeng thinks other AI models move too slowly; according to the folks I spoke with, the idea is that, most humans don’t need to read every single piece of text to interpret what action to take. Thus, AI should be the same.
Hell, Nio and Onvo’s self-driving abilities were equally impressive as Xpeng’s, able to navigate nearly to the door, and go straight to one of its battery swapping stations. While Xpeng’s tech seems good, it’s not unique—nor was the software we tested part of its AI-forward announcement of VLA 2.0 we were told about the day before.
Flying car
[https://www.youtube.com/watch?v=OKG9TGVL6x8]