"From 'pursuing performance perfection' to 'pursuing practical efficiency,' from 'humanoid obsession' to 'task fit,' from 'technical validation' to 'value realization.'"
At the 2026 Year of the Horse Spring Festival Gala, four domestic robot companies — Magic Atom, Songyan Dynamics, Unitree, and Galbot — took turns on stage. Their agile dancing, martial-arts coordination, and interactive performances flooded the internet, becoming the most talked-about tech symbols. Seeing through the phenomenon to its essence: these precise, fluid movements are essentially engineering shows of pre-programming and pre-training; today's robots still cannot understand the physical world, and lack the autonomous decision-making ability to handle the unknown.
Turing Award winner and Meta's chief AI scientist Yann LeCun punctured the bubble with a blunt remark: no humanoid robot today has genuine practical value, and their common sense is worse than even a cat's. It is Nvidia's Jensen Huang, in his 2025 GTC keynote explaining Physical AI, who points to robots' ultimate future. The humanoid direction is correct, yet the industry is currently at the critical early stage of commercial exploration; shedding the glitz and taking root in practicality is the inevitable path to maturity. And all this is merely the starting point of this long technology marathon.
01 The Truth About Performance Robots: The Result of Massive Training, the Pinnacle of Engineering Capability
Every move and every interaction of the gala robots comes from pre-calibrated trajectories and massive data-fitting training, not autonomous thinking. Songyan Dynamics' robots performed fixed interactions in a skit; Unitree's robots executed precise moves alongside martial-arts performances; Magic Atom's panda robots danced in unison to build momentum. These flawless performances are only possible in a closed, controlled stage environment. Faced with unexpected situations — stage lighting changes, an audience member wandering in, an adjusted task flow — the robots break down and cannot adapt on the fly the way humans do.
This kind of "intelligence" is typical demonstration-grade intelligence, not practical-grade intelligence. At the 2026 Davos Forum, Yann LeCun objectively noted: "All the amazing robot performances are the result of pre-programming; the industry has yet to break through the bottleneck of autonomous cognition of the physical world, and the public's expectations of robot intelligence remain biased." In terms of capability, today's robots are far below the level of an elementary-school child: a human child quickly adapts to new environments through observation and trial and error, while a robot can only repeat preset actions, with no understanding of physical rules or causal logic. The robot craze we see is a concentrated display of engineering capability, not an essential breakthrough in AI technology.
02 Physical AI: Intelligence Comes from Physical Interaction, Not Script Execution
At the 2025 GTC conference, Jensen Huang gave the official definition of Physical AI: enabling AI to move from virtual data into the physical world, with real-time perception, autonomous decision-making, and dynamic feedback — able to understand physical rules such as friction, inertia, and causality, rather than relying on preset scripts to execute actions. In the conference's closing segment, Nvidia, together with Google DeepMind and Disney, released the Newton robot physics-training engine (an open-source physics engine), and showcased Blue (the BDX-series robot), a physical robot developed through the three-way collaboration. Inspired by Star Wars, this small robot performs no flashy somersaults or dance routines, yet can adjust its movements in real time to environmental changes, avoid obstacles autonomously, and complete tasks on its own; every reaction comes from real-time inference, not pre-programming. It can recognize human speech and show corresponding emotions — "proud" when Huang praises it, "impatient" when hurried — demonstrating Physical AI's interactive nature. Intelligence is not trained; it is generated through interaction with the physical world. Physical AI pursues "understanding the world and acting autonomously," rather than mechanically replicating movements. Compared with the gala robots' "performance intelligence," Blue may look plain, yet it represents the direction of Physical AI: free from script constraints, with genuine physical adaptability and decision autonomy.
Pre-programming and pre-training are necessary technical accumulation. Without today's engineering polishing and scenario validation, there would be no comprehensive future breakthrough of Physical AI — performance is the start, productivity is the goal. Short-term scenario-based demonstrations are an important form of market education and technical validation; over the long term, focusing on underlying intelligence, core components, and practical scenarios will eventually achieve a dual breakthrough in technology and business.
03 Commercial Deployment: From Performance to Productivity
The humanoid form is the mechanical shape best suited to human production and life, with enormous future potential. The industry is currently in the critical transition from technical validation to commercial deployment, and has not yet achieved large-scale productivity conversion — a common challenge for humanoid robots worldwide.
First, costs remain high, and continuous breakthroughs are needed to reduce them at scale. Industrial-grade humanoid robots generally cost ¥500,000 to ¥2 million each, with core components accounting for over 60% of the cost; the payback period still needs optimization and is a primary consideration for enterprises' bulk procurement.
Second, shipments are growing steadily and application scenarios are gradually expanding. According to authoritative UBS data, global humanoid robot shipments in 2025 were about 18,600 units, with Chinese enterprises dominating the share; global demand in 2026 is projected to reach 30,000 units, and companies such as AgiBot, Unitree, and UBTECH continue to increase shipments, with products covering diverse scenarios like research, exhibition, and industrial testing.
Third, non-humanoid robots are achieving commercial breakthroughs first, providing a reference for the humanoid track. Unitree's quadruped robot dog has been deployed in scenarios such as hydropower-station inspection and power-grid maintenance, proving that task-fit robots are easier to commercialize early.
From "pursuing performance perfection" to "pursuing practical efficiency," from "humanoid obsession" to "task fit," from "technical validation" to "value realization." As robots move from the gala stage to factory lines, home services, and fields and farms; as pre-programming gives way to autonomous decision-making; as demonstration intelligence evolves into Physical AI — Chinese robots will surely become the core carrier of new-quality productivity.
The gala robot show is a starting point, not an end; the glitz of pre-programming is a transition, not the final state. The era of Physical AI has not yet arrived. Shed the glitz, take root in reality, and let robots move from "knowing how to perform" to "able to create," and we will eventually welcome an intelligent future co-created by humans and machines. And this technology marathon concerning the future has only just begun.
