Momenta R7 Reinforcement Learning World Model Achieves First Mass Production, Opening a New Chapter in Physical AI | Unity Ventures Portfolio Shines at Beijing Auto Show
Physical AI Moves From Technical Concept to Mass Production and Real-World Deployment
Unity Ventures portfolio company Momenta recently hosted a themed event at the Beijing International Auto Show titled "Momenta R7: The Prologue of Physical AI." On the occasion of the company's tenth anniversary, four partners took the stage together for an in-depth dialogue covering Momenta's technical philosophy, product capabilities, and global commercialization. They officially announced the first mass production launch of the Momenta R7 reinforcement learning world model — a milestone marking intelligent driving's leap from "seeing the world" to "understanding the world," and the transition of physical AI from technical concept to scaled mass production.
Understanding the Nature of Intelligence: Achieving General Intelligence Through Physical AI
At the event, Momenta partner and CEO Xudong Cao offered a detailed breakdown of physical AI's underlying logic and evolutionary path. Prediction is the foundational cornerstone of intelligent evolution: large language models rely on Next Token Prediction to compress commonsense knowledge of the digital world, giving AI the ability to understand text and natural language; world models, through World Model Prediction, predict future states and interaction logic in the physical world, gaining understanding of objects' physical properties, causal relationships in motion, and potential possibilities for interaction.
Autonomous evolution is the key driving force for continuous iteration of intelligence. AI operates within environments, continuously receiving feedback, trial-and-error iteration, and autonomous optimization based on goals. This evolutionary logic is the core essence of reinforcement learning.
Therefore, world models and reinforcement learning together constitute the two pillars of physical AI.
World Models and Reinforcement Learning: Building the Foundation of Physical AI
Momenta partner and Senior VP of R&D Yan Xia provided a deep technical architecture breakdown of the Momenta R7 reinforcement learning world model at the launch.
Xia noted that the core of physical AI lies in deep cognition of the world's fundamental physical laws. Momenta's world model operates at three levels:
Level one: world model pre-training. Through pre-training on massive amounts of real driving data, physical laws, commonsense, and causal relationships are compressed into the model, enabling the system to form foundational cognition of the physical world;
Level two: world model simulation. The world model is applied to closed-loop simulation for autonomous driving, allowing the system to extrapolate how the world will evolve when its own behavior changes. Relying on efficient scene extrapolation capabilities, it performs performance evaluation on long-tail scenarios;
Level three: reinforcement learning within the world model. Building on the first two levels, this constructs a highly realistic virtual training ground for reinforcement learning, allowing the system to repeatedly explore and trial-and-error in environments close to reality.
It is within this virtual training ground that reinforcement learning can fully realize its advantages. The system moves from simple "Imitation" to "Imagination," undergoing millions of extrapolations in the virtual world and autonomously acquiring the ability to make optimal decisions in complex game-theoretic situations. Extreme scenarios that occur once in ten thousand times in the real world are repeatedly experienced and thoroughly digested in this virtual training arena. The goal of this mechanism is to enable the model to outperform human drivers in rare extreme scenarios — and this is where the true value of physical AI lies.
One-in-Ten-Thousand Extreme Scenarios: Safeguarding with Ultimate Intelligence
Momenta partner and Senior VP of R&D Gang Sun emphasized that true physical AI is not just about enabling vehicles to drive smoothly in the vast majority of everyday scenarios, but more importantly, providing users with safe, reassuring, and ultimate intelligent protection even in one-in-ten-thousand rare extreme scenarios.
At the event, Sun presented multiple groups of real-world driving scenario demonstrations, visually showcasing the outstanding performance of the Momenta R7 reinforcement learning world model in extreme situations.
Anticipating an Apple falling "out of nowhere"
Seizing the "fleeting moment" across space and time
A "effortlessly masterful" spatial utilization expert
A "future-foreseeing" master of rational driving
Long-tail extreme situations that traditional end-to-end algorithms struggle to preset and cover are the true proving ground for physical AI's value. The Momenta R7 reinforcement learning world model's composed responses in these situations confirm its core capability: truly understanding the laws of motion and interaction logic of the physical world, rather than relying on scene memory and rule matching — this is precisely the fundamental reason why the Momenta R7 reinforcement learning world model can provide users with comprehensive reassuring protection in one-in-ten-thousand extreme scenarios.
China's Intelligent Driving Flywheel Speed: The Shared Choice of Global Brands
Momenta partner and Senior VP of Marketing and Business Huan Sun shared the company's latest progress in global commercialization at the launch.
The data speaks most powerfully. Over the past year, the scale of mass-produced vehicles equipped with Momenta's system jumped from nearly 300,000 to over 800,000, an increase of more than 160%. Mass production speed has also achieved a qualitative leap: the first 100,000 units in 2022 took 24 months, while now 100,000 units can be delivered in as fast as under 40 days. The flywheel effect is accelerating.
Currently, Momenta has successfully delivered over 70 mass-produced models, with cumulative designated models exceeding 200. At this Beijing International Auto Show, more than 20 brands and over 60 models were equipped with Momenta's intelligent assisted driving solutions, spanning a diverse matrix including domestic brands and Sino-foreign joint venture brands — the most powerful testament to "the shared choice of global brands."
In terms of global expansion, Momenta's solutions have already landed in over 10 countries and regions including the United Kingdom, Norway, Singapore, and Australia-New Zealand. Meanwhile, the Momenta Robotaxi global ecosystem continues to expand, with strategic partnerships established with multiple ride-sharing companies and OEMs to jointly advance the scaled commercialization of autonomous driving services, providing users worldwide with safe, convenient, and efficient mobility experiences.
Original Aspiration as Mission: Better AI, Better Life.
Reflecting on the founding mission, Cao recalled with emotion: ten years ago, he encountered Fairchild Drive in Silicon Valley, the street named after Fairchild Semiconductor that gave birth to the starting point of the global semiconductor industry and ultimately forged the legendary foundation of Silicon Valley. "In that moment, the fire in my heart was instantly ignited; and today, we hope to join hands with all Chinese AI companies to jointly write the Silicon Valley legend of the East."
Ten years is a starting point, and also a proof. Momenta's mission and vision remain steadfast: save a million lives in ten years, liberate 100% of time in ten years, double logistics and mobility efficiency in ten years.
The release of the Momenta R7 reinforcement learning world model is the opening of the prologue of physical AI, and also the starting point of the next decade. Momenta will continue to work with global partners to push intelligent driving to truly benefit every user through technological breakthroughs. The prologue has begun, and the journey continues unceasingly. Better AI, Better Life.
