Opening with that bold statement: China is no longer just building robots it’s choreographing them. Humanoid robots at the 2025 Understanding China Conference in Guangzhou did not simply stand as static exhibits. They moved with fluid precision, throwing calculated punches and executing dance routines that highlighted both engineering mastery and AI sophistication. Performance was more than mere entertainment, a live demonstration of China’s accelerating progress in humanoid robotics, now supported by huge state investment and a rapidly maturing supply chain.
Behind every one of those spectacular scenes of boxing robots, there is some complex integration of mechanical engineering, control theory, and AI-driven motion planning. In order for robots to perform credible boxing and dancing, finely tuned actuation systems and balance control algorithms are needed, based on ideas like ZMP regulation. The engineers will make sure that the ZMP of each foot remains inside its support polygon through dynamic adjustments of CoM, necessary for maintaining stability by means of rapid changes in posture. This becomes critical when fast and high-impacting motions-like punches-must be performed, whereby the pitch angle of the torso and contact forces of the feet need coordination on a millisecond time scale to avoid destabilization.
In various demonstrations, hierarchical whole-body control frameworks were likely used where the highest-priority task is balance, followed by foot positioning, linear momentum control, and then expressive upper-body movements. For such systems, multiple task layers are solved with quadratic optimization algorithms-running often at 1 kHz-while respecting safety constraints like joint torque saturation and friction cone limits. Integrating such expressive motions without losing balance shows great advances in multi-task control and null-space projection methods.
These humanoids represent the convergence of perception, planning, and actuation, from an AI perspective. Motion planners may bring in imitation learning, whereby robots observe human movements and replicate them through embodied AI models that fuse visual, proprioceptive, and tactile data. Such a method allows far more naturalistic motion trajectories; one can easily make the robot modify the footwork and arm movement in real time to match rhythmic patterns or even interact with simulated opponents in boxing sequences.
The country’s humanoid robotics sector has grown large-from about 100 companies worldwide at the beginning of this year to 150 in China alone. Underpinning it all is robust support from the government, with the Ministry of Industry and Information Technology putting in plans to develop a full-stack humanoid ecosystem by 2025. Firms such as Unitree have already captured headlines with sub-US$10,000 humanoids, and UBTech claimed the “world’s first mass delivery” of industrial humanoid robots. These are just part of the spate of announcements that suggest a transition from lab prototypes to scalable platforms-even if most models today remain optimized for demo rather than heavy-duty industrial work.
The Guangzhou showcase also illustrates China’s strategic focus on ’embodied AI’ – systems with the ability to perceive, reason, and act in the physical world. These robots can execute sophisticated sequences in dynamic environments autonomously, independent of human teleoperation, with multimodal sensing integrated with AI-driven control. This aligns with Beijing’s broader objective of embedding AI into the “real economy,” putting humanoids to use in manufacturing, logistics, healthcare, and even possibly defense.
Technical challenges are far from overcome: almost all humanoids have uptimes of two to four hours per charge, though solutions like swappable battery packs, or even autonomous battery changes, herald continuous operation-as demonstrated by the Walker S2 from UBTech. The other frontier is dexterity; the robotic hands of today’s machines lack the degrees of freedom and fine sensorimotor integration that enables such manipulative tasks as tying shoes or preparing small meals, which severely limits utility outside of structured tasks. And essential is cost reduction: significant unit price cuts from today’s $150,000 to $500,000 to $20,000-$50,000 for widespread adoption will require modularized mechatronic joints, easier wiring, and platform standardization.
Yet, the performances in Guangzhou suggest that at least in a controlled exhibition setting, China’s humanoids already cross the first bridge: safety in fenceless operations. Their capacity for conducting dynamic routines without incident speaks to robust real-time motion planning, compliant actuation, and effective fall mitigation strategies. As these capabilities mature, the same engineering principles that keep a robot balanced mid-spin could one day keep it stable while lifting heavy loads on a factory floor or assisting patients in a hospital ward.
In the global race for humanoid robotics, China’s mix of rapid iteration, supplier-led innovation, and state-backed scaling could prove decisive. The dance and boxing routines in Guangzhou were more than a show; they were the signal that the country’s humanoid platforms are evolving into agile, expressive, and increasingly autonomous systems, positioned to step out of the exhibition hall and into the real economy.
