
From HKEX to Munich: How Momenta Put Physical AI into a Million Cars
From HKEX to Munich: How Momenta Put Physical AI into a Million Cars
Physix FrontierKeeping investors focused on global autonomous driving and robotics leaders, tracking products, orders, and industry changes.
On November 18, if you open the Uber App in Munich to call a ride, the car that shows up might drive itself: the steering wheel turns on its own, with a safety operator in the back seat. Uber and Momenta announced that day that the two sides had launched Robotaxi trial operations in Munich, initially covering the area around the Olympic Park in the north, with free rides opened up in batches.
This company went public just four months ago. On July 8, Momenta listed on the Main Board of HKEX under the ticker 6880.HK, with an offer price of HK$295.6, raising about HK$6.8 billion. The public offering was 414 times oversubscribed, and its market cap exceeded HK$70 billion on the first day of trading. Another number announced the night before listing: production vehicles equipped with its system surpassed 1 million units. The millionth was an IM LS8, and its owner was invited on stage to ring the bell together with Cao Xudong.

At the launch event, four partners on stage together
These developments trace back to April 25. That day at the Beijing Auto Show, Momenta held a "Prologue to Physical AI" sharing session, coinciding with the company's tenth anniversary, where the four partners appeared on stage together for the first time: CEO Cao Xudong, R&D SVP Xia Yan, R&D SVP Sun Gang, and Marketing & Business SVP Sun Huan. At the event, they announced the mass-production debut of the R7 reinforcement learning world model. Cao Xudong said this means intelligent driving is moving from "seeing the world" to "understanding the world." Many talk about physical AI, but few have actually put it into production cars.
1. Starting from "Prediction"
Cao Xudong's explanation of physical AI starts with the word "prediction."
Large language models predict the next word, and by learning the common sense embedded in text, they can understand human language. World models predict the next step of the physical world: how things move, how causes connect to effects, and what happens when people and machines come into contact. In Cao Xudong's words, only when prediction is accurate can one be considered to understand the world.
Another thing is autonomous evolution. Once AI sets a goal, it tries things in the environment, corrects mistakes, and keeps trying after corrections—this is reinforcement learning. Cao Xudong combines the two: world models plus reinforcement learning underpin physical AI.
Put this way, physical AI is no empty phrase. The model must understand the laws of the world, and it must also practice within those laws on its own.
2. The Three-Layer Structure of the R7 World Model
How does this land on vehicles? Xia Yan explained the structure of the R7 world model, which has three layers.
The first layer is pre-training, feeding large amounts of real driving data to the model so it first grasps the laws and causality of the physical world. The second layer uses this world model for simulation, reasoning "if I drive like this, what happens next," and uses it to examine rare scenarios. The third layer is the most critical: running reinforcement learning directly inside the world model.
Only after the first two layers are in place does the third layer come into play. Xia Yan said the system has shifted from "imitation learning" to "imagination and exploration." In real-world driving, many extreme scenarios may never be encountered in a lifetime, but the model has already practiced them many times in the virtual training ground. His desired outcome is clear: in rare scenarios, the model should drive better than a human.
This is where R7 differs from traditional end-to-end approaches. Traditional approaches rely on seen data and hard-coded rules, and they falter when encountering unseen scenarios. R7 first thoroughly learns physical laws in the virtual world, then makes judgments.
3. Validating with One-in-Ten-Thousand Scenarios
After the technology, Sun Gang spoke about safety.
He said that to judge physical AI, don't look at how smoothly it drives day to day; look at whether it can handle one-in-ten-thousand extreme scenarios. Several real-world tests were shown on site: an apple falls from the car ahead, and the vehicle reacts instantly; navigating a narrow road, it plans the route ahead like a veteran driver.
Physix Frontier