WRC 2026 Shows China’s Robots Moving From Showpieces Toward Real Work

WRC 2026 presented an industry in transition. Held in Beijing’s Yizhuang district from August 19 to 23, the event showed robots sorting parcels, handling production workflows, preparing food and performing household tasks alongside the more familiar humanoid stunts. Its deeper message was that the next test for robotics is not whether a machine can impress an audience once, but whether it can perform useful work reliably enough to earn a place in an operating business. 3417

At the same time, the conference made clear that the software foundation for broadly capable robots is still unsettled. Companies are pursuing different answers to the same problem: how to give machines enough physical understanding, training data and adaptability to work beyond a tightly scripted task.

The shift from spectacle to operational work

The event’s theme—“Human-Robot Symbiosis, Production-Demand Convergence”—put applications and buyers closer to the center of the conversation. Exhibits included parcel sorting, mobile-phone packing, material handling, reception, retail and household chores. Reuters described the industry’s ambition as a move beyond crowd-pleasing displays toward broader commercial use, while also noting that the demonstrations did not by themselves prove durable, scalable deployment. 217

That distinction matters. A robot completing a task on a prepared exhibition floor demonstrates capability; a robot completing it repeatedly in a changing workplace demonstrates an operating product. WRC 2026 showed more evidence of the first step toward the second than of mass deployment itself.

Several exhibits illustrated the range of this push:

• X Square Robot demonstrated a unified model across logistics and household settings. Its robots were shown organizing clothes, cleaning, retrieving deliveries and drinks, watering plants and handling other domestic tasks.
• UBTECH presented coordinated humanoid workflows involving lifting, stacking, sorting and replenishing goods, while its Walker C1 was positioned for service and public-facing interaction such as reception.
• JD.com used its JoyInside AI platform on a Unitree R1 humanoid, which was shown walking a robot dog. The display highlighted JD’s role not only as a potential robot user but also as an AI-platform and ecosystem participant.
• Food, retail and service demonstrations—including automated cooking, coffee preparation and robot-operated retail experiences—made the commercial intent visible to visitors. 516

The evidence supplied for WRC 2026 does not establish specific applications by AI² Robotics or LimX Dynamics, so their roles at the event should not be overstated.

A large exhibition, but scale is not the same as maturity

The conference’s numbers were substantial. Organizers reported 373 domestic and international companies, more than 3,000 products and 311 first-launch products across the five-day event. The exhibition covered complete robots, core components, industrial cooperation and applications across multiple settings. 34

The scale is significant because it points to a broader commercialization stack rather than a collection of isolated robot makers. The show brought together robot bodies, actuators and sensors, chips, data-collection systems, foundation models, application software and prospective users. Core components such as dexterous hands, force sensors, joint modules and chips also received prominent attention. 5

But the figures should be read carefully. Hundreds of exhibitors and thousands of products measure industry activity and market ambition; they do not establish that thousands of systems are operating profitably at scale. The more useful signal was the growing connection between demonstrations and identifiable jobs: moving materials, sorting packages, replenishing inventory, serving customers and completing household chores.

The unresolved route divergence in embodied AI

The conference also highlighted a fundamental strategic question: what is the most efficient route to adaptable robot intelligence?

Route one: scale human video to learn physical behavior

Dyna Robotics’ Dyna-2 represents one approach. The company describes it as a world-action model for manipulation, pre-trained on more than one million hours of egocentric human video—roughly 170 years of continuous waking experience. Its premise is that large-scale video can provide information about how scenes change, how objects respond to contact and how people perform physical actions, reducing dependence on scarce robot-action data.

Dyna says Dyna-2 can transfer across robot platforms and adapt with only hours of additional fine-tuning. Those are company claims, however, rather than independent validation of general-purpose performance.

Route two: adapt quickly from a demonstration

Generalist’s GEN-1.5 reflects a different emphasis. The company describes it as a multimodal robot foundation model that can learn a new task from a single demonstration in seconds, without gradient updates or fine-tuning. Its system processes video along with sensor, language and proprioceptive inputs and produces action trajectories.

This approach prioritizes fast in-context adaptation: instead of retraining a robot for every new task, an operator demonstrates what to do and the model attempts to generalize the instruction to new situations. As with Dyna-2, the available evidence is primarily company-reported, so the claims should be treated as reported capabilities rather than settled benchmarks.

What remains unknown

These approaches expose the industry’s unresolved choices:

• How much useful physical knowledge can be learned from human video rather than robot interaction?
• When is real-world data more valuable than simulation or recorded demonstrations?
• Should developers prioritize larger models, world modeling, imitation learning or rapid task adaptation?
• Can a model trained on one robot transfer reliably to another body with different sensors, geometry and control limits?

The available evidence does not establish a single winning architecture or training recipe. The route divergence is not a side issue; it determines how expensive it will be to add new tasks, new environments and new robot bodies.

What the industry’s “landing year” looks like

WRC 2026 suggested a two-track strategy. Companies are pursuing constrained applications that have an obvious economic rationale—warehousing, assembly, logistics, retail and service—while continuing to invest in foundation models, data infrastructure and hardware for more general embodied intelligence. 24

That combination is pragmatic. Narrower systems can generate customer feedback and operational data now, even if they do not solve general-purpose robotics. Humanoid platforms may attract attention, but wheeled systems and specialized machines can be better suited to particular tasks; reporting from the conference noted that many robots doing productive work were rolling rather than walking.

The result is a market that is commercializing before the technical frontier is settled. WRC 2026 was therefore less a declaration that general humanoids have arrived than a test of whether robotics companies can turn impressive demonstrations into repeatable workflows—and use those deployments to keep improving while the competition over embodied-AI methods continues.