At XPENG's AI Day last November, I debuted our full-stack Physical AI framework, aiming targeting to bring Physical AI applications, including Robotaxis, humanoid robots, and flying cars into mass production by 2026.
Now, at this mid-year juncture, we are looking at the real key to scaling physical AI.
What is Physical AI
First, what is Physical AI? You might also frequently hear it referred to as 'embodied AI.'In fact, Physical AI encompasses a larger domain than embodied AI.
Integrating the AI capabilities of the digital world with physical hardware, such as autonomous vehicles and robotics, will give rise to 'Physical AI'. For example, robots will progressively acquire the capacity to comprehend, interact with, and reshape the world, ultimately triggering a revolutionary shift in productivity and the relations of production.
Physical AI encompasses four core elements
To scale Physical AI, we need to break down its core elements. In my view, Physical AI encompasses four core elements: models, computing power, data, and physical embodiments.
The bedrock of the Physical AI world is the operating system, and the model is that operating system. Meanwhile, the large model can be viewed as the engine. Data is the fuel driving its evolution, and the scale and efficiency of data application determine the capabilities of the model.
Models, computing power, and data are elements belonging to the digital realm that follow Scaling Laws in virtual space, meaning model performance continues to improve as model parameters, computing power, and dataset volume expand.
By contrast, physical embodiments fall under the physical world. They refer to tangible systems empowered by AI, such as vehicles and humanoid robots, whose capabilities are constrained by the physical laws governing manufacturing.
These four elements collectively form the foundational backbone of autonomous driving and even artificial general intelligence (AGI). Real-world deployment of Physical AI can only be achieved through simultaneous breakthroughs in both digital and physical elements.
Mass Production: Physical AI vs. Digital AI
Without a doubt, mass-producing Physical AI is vastly more challenging than digital AI. But beyond software constraints, what other factors are at play？
- Information density：Digital AI deals with low-density information streams. Physical AI handles much higher-density information streams. Therefore, the transition from digital AI to Physical AI represents a shift from low-density information streams to high-density information streams, as well as a transition out of the digital domain into physical space-time.
- Capability boundaries：The upper limit of digital AI is found in higher information efficiency, featuring fault tolerance at its lower bound and highly transferable applicability. Conversely, the upper limit of Physical AI is its power to reshape the physical world, while its lower bound dictates strict safety standards with zero tolerance for error, and its applicability remains deeply case-specific. Crucially, digital AI is universal and easily transferable, but Physical AI is deeply tied to specific scenarios.
- Hardware barriers: While CPU, GPUs, and server clusters constitute the primary hardware barriers for Digital AI, the hardware barriers of Physical AI cover far broader dimensions, including the fundamental performance, cost, reliability, manufacturing capacity, and mass-producibility of edge-side hardware.
- Laws and regulations：Digital AI regulations focus on indirect management, involving data privacy, copyright, and ethics. By contrast, Physical AI faces direct and strict operational restrictions imposed by policies and regulations. For example, Robotaxis require road-testing permits and rigorous safety certifications.
- Public acceptance: Digital AI boasts prominent tool traits and wins easy public recognition. Physical AI, however, involves safety and credibility concerns and requires a much longer cycle to foster public trust.
This is why the race to scale Physical AI belongs to companies with cross-domain integration and self-development capabilities, not just those who focus solely on models or hardware.
