Over the past few years, big models have pushed AI's capabilities into the digital world. But as AI moves from text and code to the real world, more complex questions have begun to arise: how do machines understand the world? how do they predict change? how do they continue to operate in the real world?
This is why Physical AI (physical intelligence) is becoming one of the hottest racetracks in the global tech industry.
Recently, Zhang Yi, the head of an end-to-end core technology division of one of the world's leading autonomous driving companies, announced his departure to focus on general physical AGI.
However, they are one of the first in the country to explicitly explore "AI for Physical AI", and are committed to achieving true AI for Physical AI through Auto Research. Their long-term vision is for North American frontline companies such as Standard Generalist to improve the model's ability to generalize zero-sample and few-sample in new tasks, new scenarios, and drive universal physical intelligence to commercial applications.
The core idea is to open up automated research systems with data, train, evaluate and deploy infrastructure to expose real-world execution issues, continuously transform them into new data, evaluate and research tasks, and drive model transformation. The team believes that models determine the limits of capacity, and Auto Research and Infra determine the speed at which capacity evolves.
As the country's most representative real-world technology leader in the field of autonomous driving, during his tenure as chairman, Zhang was one of the youngest technical leaders in the P10 class. Today, he chooses to extend his end-to-end capabilities through real-world scenarios and large-scale product verification to a physically intelligent world that can learn and evolve autonomously.
It is understood that the company was founded in July this year and has completed several rounds of financing totaling hundreds of millions of yuan.
Behind the rapid accumulation of capital, a trend is becoming clear: as Physical AI moves from conceptual exploration to industrial deployment, the industry needs not just single-point algorithm breakthroughs, but the ability to integrate models, data, engineering systems and real-world scenarios.
Zhang Yi's start-up brought not only a set of AI mass production experiences that have already been verified in the process of self-driving mass production, but also introduced Auto Research to allow AI to continuously evolve into a whole new R&D model, the Physical AI pathway, introducing new gameplay.
The "end-to-end mass production" capabilities practiced in "real war"
In the AI industry, there is always a "gap" between algorithmic capabilities and industrial settings: a model that performs well in an experimental environment, demo cool, doesn't mean it can run reliably in the real world.
Autonomous driving, in particular, requires vehicles to cope with different weather conditions, complex traffic actors, long-tail scenarios and a large number of unpredictable changes. Models not only need to "understand" but also need to make decisions steadily in a continuously changing environment.
Chang'an's engineering background has been built in the process of facing these real challenges.
It is understood that Zhang Chen initially switched from perceptual technology to being responsible for the direction of dynamic barriers, leading the team to achieve a transversal improvement of critical capabilities for more than six months. He then gradually grew from the base of the algorithm to the head of core technology. This experience not only helped him understand how to improve model capabilities, but also gave him mastery of the various engineering challenges that algorithms must overcome after entering the real physical environment.
At the time, the "end-to-end" concept swept through the autonomous driving industry and was seen as the inevitable path to high-level physical intelligence, but most teams remained in the model verification or small-scale road testing phase.
This is because the difficulty lies not only in the design of the algorithm architecture, but also in how to keep the model stable in real-life highway scenarios, create efficient data locks, and implement engineering landings with limited computing and end-to-end resources.
This means that while the industry is still discussing the end-to-end theoretical framework, Zhang has completed the entire cycle from technical exploration, organizational aggression, engineering verification to mass production landings.
This mass-production closed-loop experience "from 0 to 1 plus to N" is rare in the domestic end-to-end field.
In the eyes of the insiders, he was able to stand alone, while raising the technical direction and the new backbone of the team.
This experience also explains why Zhang Yi chose the next stop in Physical AI. Autonomous driving is essentially a special landform of physical intelligence in the road traffic scene, both facing a highly similar underlying proposition: how to make machines understand the environment reliably, reasonably predict the future, and continue to make accurate action decisions in uncertain physical spaces.
Chang's previously verified end-to-end architectures, data locks and systems engineering methods provide a solid engineering foundation for the current physical intelligence shifting from algorithmic verification to reliable deployment.
Top puzzles: academic, engineering and machine-controlled complementary models
Over the past few years, the core of AI competition has been large model capabilities, while entering the phase of Physical AI naturally requires stronger interdisciplinary collaboration capabilities.
In other words, its technical links naturally span multiple domains: it requires both basic models and world models to provide cognitive capabilities, it requires robotic control systems to perform real actions, and it requires engineering teams to integrate these capabilities into real-world scenarios.
Around this direction, Zhang quickly assembled a tight-fitting model that fully complemented academic, engineering and real-world landings.
He is currently an assistant professor at the Hong Kong University of Science and Technology and the head of the World Mind Lab, a world leading expert in the field of models and physical intelligence. Prior to this, he worked for a long time at MIT and Harvard University, focusing on 3D world modeling, research covering 3D vision, generation models, neural imaging, 3D reconstruction and physical robots, and published more than 50 highly cited papers in top international academic conferences and journals such as SIGGRAPH, CVPR, ICCV, TAMI.
In the underlying architecture of Physical AI, robots must have the ability to predict and simulate three-dimensional physical space, building a world model with a deep understanding of the laws of physics. The addition of cubic energy, so to speak, builds the academic foundation for the team's exploration of the underlying spatial representation and the forefront of world models.
In the realm of machine control and robotics, the team brought in a young scientist with both an international academic perspective and deep industry background experience, Liu Xiaobo.
With extensive expertise in robotic learning, three-dimensional perception, robotic operation, and human-robot full-body control, Xi Mazeyu's research is one of the industry's first representative achievements in the field of universal human-robot controller authentication.
He was a founding member of Assured Robot Intelligence (ARI), which was officially acquired by Meta and merged into Meta Superintelligence Labs.
Today, Xi chose to return to China to start a business together with Zhang Yi, focusing on building a new generation of models with a physical intelligence base capable of autonomous learning and continuous evolution.
To this end, the three core members, each with a deep technical longboard, have put together a system of capabilities that covers the entire chain:
Zhang Yi: Manage the end-to-end mass-production cycle with the integration of complex systems engineering to solve "how to really make smart systems and push them to real applications";
Accountability: Provides underlying 3D visual reconstruction with a world-modelling front-end academic engine to solve "How machines understand the real world";
Xi Mazeyu: Open human-robot full-body dynamic control, agile operation and real-world landing verification, solving "how machines perform tasks in the real world".
The academic frontier, engineering systems, and entity-controlled triangles are closely intertwined to form a "Golden Triangle", a whole chain from "Basic World Model Construction to Silicon, Complex Model System Engineering to Silicon, Physical Equipment to Control Landing".
It's the "scaled moment" of physical intelligence.
The team has completed several rounds of financing in just one month, amounting to hundreds of millions of yuan, which reflects the recognition of the team by top capital from the side.
In the past, the market has often focused on the founding team's academic background and laboratory outcomes. However, as IDC has recently emphasized in relevant industry insights, Physical AI is currently entering a critical phase in the transition from technical competitive to industrialized competitive capability. As the industry as a whole moves gradually from concept testing to product verification and real-world scenario verification, investors and industry are becoming increasingly aware that the long-term competition of Physical AI is not an isolated struggle of single algorithms, but an extremely challenging system integration, engineering efficiency and "comprehensive run-off" of data pilots.
The Ministry of Strategic Investments of Xiaomi said: "Changzhou and the team have a deep accumulation of autonomous driving end-to-end production landings, moving it to the Physical AI platform that allows the front-end model to truly run the scenic loop, which is the most critical part of the move towards industrialization with physical intelligence. Looking forward to further deep collaboration".
In the future, Zhangjiajie and his team are looking forward to starting with their own intelligence, expanding further into the broader field of Physical AI, driving physical intelligence from single-point application to building basic model and systematic capabilities".
This also indicates that the scarcity of "real warfare" in Physical AI, which is moving from "concept verification" to "scale deployment", will undoubtedly win the favor of capital.
From autonomous driving to Physical AI, if Zhang Yi's previous professional proposition was to solve "how to get end-to-end AI into mass production", then the next question to be answered is "how to achieve true AI for Physical AI using Auto Research".
So, for the team that just finished their debut, "The Hard War" is just getting started...
It is understood that the technical approaches and system transformation within the Zhang Liu team are progressing steadily. Automation research tools have entered the pilot opening phase, and model capabilities have reached industry-leading levels on some indicators. Soon, the team will officially disclose the results of specific models and product progress at the appropriate time.