Autonomous driving is the first real-world test for the AI models that will eventually power everything from robots to urban traffic systems, according to Momenta CEO Cao Xudong.
(Beijing) — Autonomous driving software provider Momenta is positioning itself to be the OpenAI of the physical world, launching a new world model that uses data from more than 800,000 vehicles to predict real-world physics and driver behavior. The company’s R7 model, unveiled at the 2026 Beijing Auto Show, is the foundation for what Momenta calls “Physical AI,” a technology it believes will extend from passenger cars to logistics and trucking.
“Autonomous driving has entered a stage where it can achieve a positive feedback loop between data and commercialization,” Cao Xudong, CEO of Momenta, said in an interview. “It is the prologue to Physical AI because it is the first scaled application to solve the data acquisition and business model problems that have held back robotics.”
The R7 world model works in three layers: it first pre-trains on massive real-world driving data to understand physical laws and causal relationships, then uses simulation to predict how the world evolves based on different actions, and finally uses reinforcement learning to train the system to make optimal decisions. This structure is designed to create a driving brain that learns from the collective experience of its fleet, which includes over 70 models from automakers like Mercedes-Benz, Audi, and BMW.
The strategy requires immense capital, with Cao estimating that achieving scalable Level 4 autonomy will require at least $10 billion in investment. By securing mass-production contracts for its driver-assistance systems, Momenta has built a cash-flow business to fund the development of a unified AI model for all vehicle types, a key advantage over competitors solely reliant on venture funding.
The core of Momenta’s strategy is creating a data flywheel that is scarce in the physical world. While digital AI exploded thanks to vast, low-cost text and image data from the internet, training AI for physical tasks like grasping a cup or navigating a construction zone requires expensive hardware and real-world interaction. Autonomous vehicles, however, are mobile sensor platforms that continuously collect complex data, solving the data scarcity problem.
“Every one of the 800,000 vehicles we have on the road is a data collection node,” Cao explained. “OpenAI’s models evolve with user queries; our models evolve with every mile driven in the real world.”
This data feeds the R7 world model, which moves beyond simple imitation of human drivers. According to Cao, raw data contains both good and bad driving habits. After pre-training to learn the "common sense" of driving, the model undergoes a post-training phase, similar to reinforcement learning from human feedback (RLHF) in language models, to align its behavior with that of a skilled, safe driver, not an average one.
Momenta’s ambition extends beyond being a Tier 1 supplier for passenger cars. The company is already applying its unified model to Robovan logistics vehicles and plans to enter the Robotruck market next year. The underlying belief is that a single, powerful driving model can be adapted to any vehicle type, creating a platform advantage similar to what platform companies achieved in e-commerce.
“We believe one autonomous driving large model can realize all autonomous driving vertical applications and do it better,” Cao said. This approach lowers the development cost for each new vehicle type while the data from each vertical—be it taxi, logistics, or trucking—improves the core model for everyone.
This platform strategy is Momenta’s answer to the immense cost of entry. Cao estimates that developing a general-purpose robot could require investments in the hundreds of billions of dollars. Without a self-sustaining cash-flow business, such a venture is unrealistic. Momenta’s success in the assisted driving market, with over 200 vehicle models signed, provides the financial engine to pursue the much larger prize of full autonomy.
This article is for informational purposes only and does not constitute investment advice.
