When Robots Leave the Demo Zone: How China’s “use-first” approach is quietly reshaping physical AI
A Series of Papers on comparison of AI in China and in Europe
Read time: ~8–9 minutes
TL; DR: China isn’t just building robots—it’s letting them run at massive scale in the real world, creating a feedback loop of data, learning, and influence that Europe hasn’t matched.
In my previous article on AI innovation in the Greater Bay Area, I mentioned a recurring impression: China’s technological evolution rarely lingers at the “proof of concept” stage. Instead, China’s strategy embeds AI within industry, enters real social systems as early as possible, uses large-scale operation to force the maturation of engineering, business models, and institutions.
That impression was amplified again this year while looking at the humanoid robot zones at CES. Judging by the distribution of exhibitors, Chinese companies have formed a highly complete industrial spectrum—from robot bodies, joints, motors, and dexterous hands to full system integration (in the chart below, Chinese firms accounted for close to 60%). Many of the exhibits felt less like speculative “future visions” and more like systems that had already been tested in real-world scenarios, now refined and systematically packaged for display.
But the real point of differentiation lies in what happens after these technologies return to society:
Who uses them?
How do they scale?
And do they ultimately become part of the infrastructure layer?
Drones: From Consumer Electronics to Urban Operating Nodes
China
In China, drones have formed a three-layer integration across consumer use, public systems, and logistics networks.
At the consumer layer, DJI-grade aerial cameras have become standard tools for creators and tourists alike. Aerial footage is now a foundational visual language on short-video platforms.
At the public governance layer, traffic police, urban management, and emergency-response departments routinely deploy drones for patrol, evidence collection, and incident response. In cities such as Shenzhen and Guangzhou, highway police operate dedicated drone units as part of their formal staffing.
At the logistics layer, drones are used to deliver everything from takeout and groceries (Meituan, Yonghui) to emergency medical supplies (Antwork) and even tourist goods at landmarks like the Great Wall. Nationwide drone delivery volumes had already reached the million-parcel scale.
UK / Europe: regulated airspace users
In the UK and Europe, drones are also widely used—for aerial photography, surveying, agriculture, policing, and infrastructure inspection. However, their mode of integration looks different:
Professionalized systems centered on licensed pilots, airspace permissions, and mission approvals
Centralized deployments led by police, search-and-rescue, or surveying agencies
Commercial delivery and dense urban airspace use largely confined to pilots and regulatory sandboxes
Technical capabilities are broadly comparable. The difference is conceptual: drones are more often treated as airspace users requiring strict institutional control, rather than as foundational nodes that can be rapidly embedded into urban operating systems.
Autonomous Delivery Vehicles
China
In cities such as Beijing, Shanghai, Shenzhen, Hangzhou, and Guangzhou, platforms including Meituan, JD.com, and Cainiao have already integrated low-speed autonomous delivery vehicles into their last-mile logistics networks.
These vehicles operate on fixed routes across campuses, industrial parks, and residential communities. Nighttime autonomous delivery has become a normalized supplement to human labor. China Low Speed Automated Driving Industry Alliance (LSAD) and the New Strategy Low Speed Automated Driving Industry Research Institute predicted that by 2026, shipments are expected to climb to 150,000 units, and are projected to reach 750,000 to 1.05 million units by 2030, forming a market space worth tens of billions of yuan. Crucially, these vehicles are not treated as showcase pilots. They are already recognized as a formal capacity type within the real dispatch systems of e-commerce and food delivery platforms.
UK / Europe
In Europe, autonomous delivery robots and vehicles do exist—Starship, Amazon Scout, and others—but their deployment is typically limited to:
University campuses
New-town or smart-city pilot zones
Regulatory sandbox cities
Their scale remains constrained, and they have yet to be deeply embedded into the core operational networks of national e-commerce or food delivery platforms.
Robot-taxis
China
Robotaxi deployment in China has moved beyond experimental pilots into meaningful commercial scale, with multiple companies operating fleets of significant size across urban markets. By late 2025:
Leading autonomous mobility companies—including Pony.ai, WeRide, and Baidu’s Apollo Go—each operate robotaxi fleets numbering over 1,000 vehicles in aggregate, pushing the Chinese industry into the “thousand-vehicle” era.
WeRide’s global fleet has surpassed 1,000 robotaxis, actively providing fully driverless services in cities such as Guangzhou and Beijing.
Pony.ai’s fleet reached around 1,159 vehicles by the end of 2025, exceeding its annual expansion target and beginning full commercial operations of next-generation robotaxis in multiple megacities.
Baidu’s Apollo Go has delivered millions of rides, with a fleet deployed across more than a dozen cities and continuing to grow.
This places China’s autonomous taxi systems at a scale rarely seen outside of a handful of early adopter regions worldwide. Instead of isolated demo corridors, these robotaxis are being used regularly by passengers and integrated into ride-hailing platforms, contributing to real traffic volumes and operational learning.
UK / Europe
In Europe, autonomous driving research is world-class. But its social embedding more often takes the form of:
Designated test corridors
Closed or semi-closed operational environments
Regulation-led trial programs
The technology is still treated as a system under continuous validation, rather than as a formal, platform-level source of transport capacity.
Service Robots
China
In restaurant chains such as Haidilao, hotel groups including Huazhu, Atour, and Marriott China, as well as in large shopping malls, airports, and metro systems, robots have become part of the standard operating environment.
Food-delivery robots, tray-collection robots, hotel item-delivery robots, and autonomous cleaning robots are widely deployed—often through RaaS (Robotics-as-a-Service) models. Operationally, they are treated much like elevators or automatic doors: infrastructure components embedded into daily workflows.
UK / Europe
In Europe, similar robots are more commonly found in hospitals, airports, high-end hotels, or smart-campus demonstration projects. Deployment tends to follow a “specialized optimization” logic, rather than being broadly accepted as part of the service sector’s baseline infrastructure.
Exoskeletons: From Research Equipment to Public Services
China
Over the past two years, elderly care institutions and rehabilitation centers have begun adopting exoskeletons for walking and stair-climbing assistance. Tourist attractions such as Mount Tai, Huangshan, and Zhangjiajie have piloted exoskeleton hiking rentals. Firefighting and emergency-response units use exoskeletons for load-bearing and rescue training.
Exoskeletons are starting to move from laboratory and industrial settings into everyday service systems.
UK / Europe
In Europe, exoskeleton use remains concentrated in industrial ergonomics, medical rehabilitation, and military or occupational protection. Expansion into elderly care and mass public services proceeds cautiously, typically via clinical pathways and safety certification regimes.
Why Narrative and Reality Diverge
For much of the past decade, dominant innovation narratives have portrayed Europe and U.S. as having earlier industrialization, more mature engineering traditions, more open academic systems, and more comprehensive institutional frameworks.
By intuition, such societies should be faster at pushing robotics and automation into everyday, large-scale use.
Yet reality reveals a different structural divergence.
Technical capability is not the dividing line.
The deeper difference lies in whether new technologies enter society primarily as risk objects to be rigorously validated, or as productive factors that can be embedded early.
In the UK and Europe, technologies tend to enter society by first fitting into legal and liability frameworks, diffusing gradually through pilots, ethical reviews, and public consultation, and scaling slowly within institutional boundaries.
In China, technologies more often enter high-frequency systems first—mobility, logistics, public safety, services, elder care—expose problems through real-world operation, and then retroactively shape regulation and standards.
This is not about one system being more open or more conservative. It reflects two different modes of socio-technical coupling:
One prioritizes institutions first, with technology gradually permeating.
The other prioritizes use first, with institutions evolving alongside deployment.
The Hidden Compounding Advantage
There is also a less visible—but increasingly decisive—consequence of this “use-first” approach.
Early, large-scale deployment generates something far more valuable than demos or benchmarks: continuous, real-world, embodied data. Autonomous vehicles, drones, service robots, and exoskeletons operating in open environments produce dense streams of multimodal data—vision, motion, force, failures, edge cases, and human interaction—that are extraordinarily difficult to simulate.
This data feeds directly into the training of world models and physical AI systems—models that must learn not only perception, but causality, constraints, and long-horizon interaction with the physical world.
Real deployment creates a powerful positive feedback loop:
Deployment generates real-world data
Data improves models and control systems
Improved systems justify broader deployment
Broader deployment accelerates data accumulation
Over time, this loop compounds. Engineering improves faster, rare edge cases surface earlier, unit costs fall through scale, and institutions adapt around systems that are already operational rather than hypothetical.
This is one of China’s most underappreciated advantages in robotics and physical AI—not because the technology is inherently superior at the outset, but because it is allowed to learn in the real world, at scale, under real constraints.
Importantly, this advantage is not accidental. China is increasingly aware that operational data itself is a strategic asset. This awareness is already shaping a tighter stance on data governance and cross-border data flows, including increased scrutiny of acquisitions, partnerships, and the export of large-scale real-world datasets generated by AI and robotics systems—the Manus acquisition being a recent illustrative example.
In future posts, I’ll go deeper into this dynamic: how real-world deployment reshapes world model training, why data density and feedback loops matter more than isolated algorithmic breakthroughs, and how China’s growing recognition of data as a strategic resource is beginning to influence regulation, industrial policy, and global competition in physical AI.
From AI industries in the Greater Bay Area, to the dense presence of Chinese robotics firms at CES, to autonomous vehicles, drones, service robots, and exoskeletons already operating in cities, all of this points to the same underlying question:
Whether technology truly becomes part of society depends on whether a society is willing to let imperfect systems enter everyday life early—and to treat the real world itself as the largest testbed.