On July 17, at the main forum of the 2026 World Artificial Intelligence Conference and High-Level Meeting on Global AI Governance, Su Hao, a distinguished Haoqing Professor at Fudan University and the founding dean of the Fudan Institute of General Physical Intelligence, said: “Physical intelligence is not here to replace people—it’s here to work alongside them. Its mission is to give humanity back to humans.”
“Today’s AI can fluently describe that ‘a cup will break when it falls on the floor,’ but it has never once felt the weight of a cup in its ‘hands.’ It can write code, create PowerPoint slides, and generate lengthy essays, yet it can’t help an elderly person in bed turn over even once,” Su Hao pointed out. He highlighted a core pain point in current AI development: “There’s no shortcut from illusion to reality. The only way is through humility before the unknown—step by step, climbing the ladder of cognition, one rung at a time.”
“So why do even the smartest large language models suffer from hallucinations? Because language has always been just a projection of the world—never the world itself.”
Su explained that humans have spent millions of years tumbling through the physical world—falling down, touching hot cups, and dropping balls we tried to catch. All those experiences got compressed into language over time. But large language models, from the moment they’re born, only learn this “shadow” of reality. Their hallucination problem stems from a “cognition without a body.” The only way out is to submit yourself to the judgment of the physical world: make a prediction, take an action, get corrected by reality, and then adjust your next move. This cycle of trial and error is what we call experimentation—and it marks the true starting point of physical intelligence. You simply can’t build it by just “browsing webpages” in a server room.
The world humans perceive is far larger than what language can capture. A vast amount of tacit knowledge—involving touch, real-world manipulation, and raw intuition—has never been systematically recorded. Gathering these scattered videos, physical data, and human experiences to fill in the missing rungs of the cognitive ladder is a hurdle modern science and engineering must overcome.
When it comes to the role and core mission of physical intelligence, Su Hao made it clear: “Physical intelligence isn’t here to replace people—it’s here to collaborate with us. It fills the gaps where human hands fall short. It takes over the heavy lifting—like turning a patient over or moving objects—so that caregivers can spend that time on companionship and genuine care. It handles dangerous jobs, letting humans step back behind the safety line to focus on judgment and creativity. Handing risky tasks to machines isn’t a cold-hearted efficiency play; it’s the highest form of humanism. The mission of physical intelligence is to give humanity back to humans.” He added that this revolution won’t happen overnight with one big product launch. “Think of it more like electrification: it first lights up factories and warehouses, then moves into stores and hospitals, and finally reaches every home. The value of physical intelligence doesn’t have to wait until the end—it’s already being released along the way.”
Based on this logic, Su Hao offered three key predictions for the industry:
First, the real breakthrough for physical intelligence won’t come from just iterating on model architectures—it will come from aggregating knowledge across institutional boundaries. “This aggregation is something no single organization can pull off alone. Videos live on the internet, equations are in textbooks, force-feedback data sits in various labs, and operational intuition resides in millions of workers. No one can gather all these pieces by themselves. It requires the whole industry—even society at large—to collaborate: co-build the data, co-define the standards, and co-share the simulation and evaluation infrastructure. Once that knowledge is truly fused into a single model, the physical world will have its own ‘internet moment.’ The so-called GPT moment will just be a byproduct.”
Second, the industry’s focus will shift from “how impressive the demo looks” to “how reliably it works.” “In engineering, there’s this concept called ‘nines.’ Going from 99% to 99.9%—each extra ‘nine’ is exponentially harder. The gap between a demo and a real product is stuck in those last few ‘nines.’ Generalization is the goal, but reliability is the starting point. Teams willing to grind away at those ‘nines’ will go the furthest. The most treacherous part of this road is that reliability chasm between a flashy demo and a dependable product.”
Third, physical intelligence will transform AI from a “reader” of scientific literature into a “creator” of new knowledge. “Today’s AI has read almost every human paper out there, but it has rarely, if ever, conducted an experiment of its own. Yet new knowledge is born precisely from experiments. When AI gains hands that can sense, manipulate, and verify the real world, it can formulate its own hypotheses, run its own experiments, and keep refining them around the clock. The discovery of new materials and new drugs could accelerate by several orders of magnitude.”
Su Hao concluded: “The physical world is intelligence’s oldest teacher, its most honest examiner, and the ultimate foundation for all intelligence. Our mission, as the AI generation, is to hand this answer sheet back to the physical world for grading.”
