Chongqing - In the previous article, we explored the idea of “Physical AI”—why some call it the “second half of AI,” a shift from being “smart with language” to becoming “smart with the body.” But concepts alone are not enough. In reality, three core questions must be answered to truly enable AI to understand the physical world: What kind of training environments does it need? Can knowledge be transferred from self-driving cars to humanoid robots? And in this global race, how do China and the U.S. differ in their approaches?

This time, we follow Professor Zhan Zhenfei—faculty at Chongqing Jiaotong University’s School of Mechanical & Vehicle Engineering, postdoctoral advisor, and former Ford North America data scientist and R&D engineer—into a deeper unpacking of these questions. From stove-top sensors in the kitchen to decision systems on the street, from U.S. simulation platforms to China’s real-world data, Physical AI is about re-modeling and re-operating the real world itself.

Professor Zhan Zhenfei of the School of Mechanical, Electrical, and Vehicle Engineering at Chongqing Jiaotong University is a postdoctoral supervisor, former data scientist, and R&D engineer at Ford North America. 

So, let’s step into the exploratory game of moving from large models to small actions and see what it takes for AI to not just “write code " but also cook in the kitchen and drive on the streets.

How does its training environment differ from traditional AI?

According to Prof. Zhan, traditional AI training relies on pre-processed datasets, which are effective for single, discrete tasks. By contrast, Physical AI must handle continuous, time-varying signals and simulate physical environments with high precision to predict and adapt to real-world dynamics.

A simple analogy: traditional AI is like memorizing a recipe book. Show it 100,000 polished photos of apples, labeled as “apple,” and it learns to recognize apples in new images. The process is static, clean, and detached.

But Physical AI is like actually cooking in the kitchen. To learn how to fry an egg, photos won’t cut it—you need the stove, the sensors, and the live feedback. The AI must monitor temperature, sound, vibration, and even smell, while adjusting the spatula’s angle by 3° at the exact moment the egg needs flipping.

In short, traditional AI is like “learning by looking at pictures in a library.” Physical AI “learns by getting its hands dirty—in the kitchen, factory, and road.”

Can knowledge be transferred from self-driving to humanoid robots?

Self-driving is one of the earliest commercial forms of embodied intelligence. They already satisfy some fundamental requirements: they possess a physical body and rely on multi-modal perception, including vision, sound, and force. However, their capability to act on the environment and effectively perceive the resulting feedback remains underdeveloped compared to fully realized robotic systems.

Can the experience gained from cars be transferred to humanoid robots? Prof. Zhan suggests that Physical AI follows a pattern of “cross-form transferability—but with readaption required.”

Put simply, can a robot borrow the “skills” a car has learned? Yes—but not via copy-paste. These skills must first be translated.

· Step 1: Identify invariants. Whether it’s a car or a robot, the laws of physics remain constant. Principles like “don’t slam the brakes on slippery roads” are universal—forming the physics-based “traffic rules.”

· Step 2: Build a reflex library. Just as humans instinctively withdraw their hands from heat, robots need millisecond-level reflexes—slightly crouching on slips or swinging an arm when balance falters.

· Step 3: Flexible translation. What a car encodes as “brake force = 30%” can translate to “ankle torque = 5 Nm” in a robot. Automated tools can perform these conversions far more efficiently than engineers relying on trial and error.

In short: A car’s “anti-skid” knowledge can be reused—but only after being unpacked, remapped, and adapted to the robot’s body. It’s like converting a driver’s license from cars to motorcycles: the underlying principles remain the same, but the operation must be relearned.

How do China and the U.S. differ in development?

On the topic of the widely discussed gap in Physical AI development between the U.S., China, and Europe, Prof. Zhan offered a succinct summary: the global landscape currently shows a pattern of “infrastructure led by the U.S., application deployment led by China.” In other words, the U.S. lays the foundation, China builds the house, and Europe is still drawing the blueprints.

According to Prof. Zhan, the U.S. holds the hardest “bricks”—large-scale simulation platforms like NVIDIA’s Omniverse, which essentially pave a high-speed highway for Physical AI. Whether self-driving cars, robots, or robotic arms, every system must be tested and run on this infrastructure first. The technology barrier is high, costs are steep, and the U.S. started early, so it currently leads.

China, in contrast, excels at turning foundational research into real-world applications. In autonomous driving, companies like WeRide and Pony.ai put cars directly on the streets, feeding massive amounts of Chinese traffic data into their models, which improves with every mile driven. Simply put, the U.S. figures out “how to simulate raindrops,” while China focuses on “making cars stay on course in a torrential downpour.”

Europe? It’s busy drafting standards and writing regulations. Fewer deployment projects exist, and the supply chain is fragmented. Goldman Sachs estimates that 13 million humanoid robots will be sold globally in the future; however, Europe hasn’t even fully built its production lines yet, clearly lagging.

Today’s competition is clear: the U.S. continues to strengthen its underlying technology, while China expands its application scenarios, and both are competing to be the first to secure both the “foundation and the house.”

Closing: The Paradigm Shift of Physical AI

The evolution of Physical AI has never been as simple as “switching to a new model.” It represents a new paradigm: using dynamic perception to understand the real world and leveraging action feedback to reconstruct cognitive boundaries. From large models to small motions, from simulation to real-world execution, the ultimate goal of Physical AI is not to “be more human,” but to “understand physics better.”

Those who master this path will have a ticket to the next technological revolution. In this race that spans models, hardware, perception, and theory, speed, scale, and deployability are all indispensable. The future belongs to the AI that can both write code and confidently get its hands dirty.