Chinese Automakers Are Building Robots, but Real Work Is the Test

Chinese automakers brought robots into the spotlight at Beijing’s 2026 World Robot Conference, but the real contest starts after the demonstrations end.

Chery-backed Aimoga presented machines working across retail, traffic management, healthcare, and public-service scenarios. GAC displayed robots derived from its autonomous-driving research. BYD also introduced its first humanoid for routine customer service at its Di Space experience center.

The crossover is no longer an isolated experiment. Automakers are treating embodied AI, software that perceives and acts through a physical machine, as an extension of vehicle intelligence. Their manufacturing plants, supplier networks, and service centers also give them something many robotics startups lack: controlled places to test machines repeatedly.

That advantage does not settle the competition. Specialized robotics companies such as Unitree have spent years refining locomotion, balance, and robot hardware. Automakers must show that vehicle engineering transfers to machines operating beside people in less predictable environments.

The emerging contest is therefore not automakers against one another. It is manufacturing scale against robotics specialization. China’s car companies can build complex products efficiently, but they still need evidence that their robots can perform useful work reliably.

Automakers Took Center Stage at Beijing’s Robot Conference

The important change was not the number of robots on display, but the growing presence of companies built around cars.

The 2026 World Robot Conference ran from August 19 through August 23 in Beijing’s Economic-Technological Development Area. Organizers structured the five-day program around product releases, procurement, development, and public demonstrations.

Around 3,000 products appeared at the conference, according to an event account from Beijing. Exhibits included humanoids, quadrupeds, robotic arms, companion machines, and specialized industrial systems.

Many demonstrations still emphasized spectacle. Robots danced, boxed, folded fabric, played table tennis, and interacted with visitors. Those displays made physical progress visible, especially in balance and coordinated movement.

Chery’s Aimoga booth pursued a more operational message. Instead of presenting one humanoid as a universal machine, the company organized demonstrations around six defined scenarios.

Its systems included a traffic-management robot, a household companion, a medical guidance robot, and a robotic guide. Aimoga also showed a quadruped carrying objects and following people autonomously.

The official scenario program described practical workflows rather than isolated movements. The medical robot, for example, was presented as a system that could identify visitors, support preliminary intake, and guide them through a facility.

That does not prove every workflow operates without human supervision. Conference demonstrations usually occur within prepared spaces, with known objects, controlled routes, and staff nearby.

However, the shift in framing matters. The machines were presented as service infrastructure, not merely as research platforms or entertainment products.

Aimoga also displayed live feeds of robots working outside the exhibition hall. Reported deployments included traffic assistance in Jiangyin, museum guidance in Wuhu, and customer reception at a Malaysian vehicle showroom.

These settings reveal Chery’s broader strategy. A car company already operates factories, dealerships, offices, logistics facilities, and customer-service locations. Each site can become a bounded testing environment.

GAC brought another version of the same strategy. The automaker began assembling a humanoid robotics research team in 2022, according to its conference profile.

Its earlier robot portfolio included GoMate, a humanoid platform, and the GoSide and GoMove service machines. The systems combine perception, navigation, mobility, and human interaction in different physical formats.

GAC says its robotics software builds on autonomous-driving research. That includes visual perception, spatial mapping, planning, and systems for interpreting multimodal instructions.

BYD’s entrance broadened the trend. The automaker said its first humanoid would interact with visitors at Di Space centers, which showcase the company’s vehicles and technology.

A customer-service role appears modest beside factory automation. Yet it gives BYD a familiar environment where staff can monitor the machine, collect interaction data, and measure failures.

The conference therefore captured a strategic migration. Vehicle companies are turning dealerships and factories into robot laboratories while converting automotive research into reusable components.

Why Vehicle Intelligence Transfers to Embodied AI

Modern electric vehicles and mobile robots share a surprising amount of engineering, even when their bodies look completely different.

An intelligent vehicle must perceive its surroundings, locate itself, predict motion, plan a path, and control physical hardware. A mobile robot faces the same sequence, although it has more ways to move and fail.

Cameras, depth sensors, inertial measurement units, and other devices generate raw observations. Perception software converts those signals into representations of objects, people, surfaces, and free space.

Planning software then selects an action. A vehicle might change lanes or stop. A service robot might approach a visitor, avoid a chair, manipulate an object, or ask for clarification.

The final step is physical control. Vehicles translate commands into steering, acceleration, and braking. Humanoids coordinate joints, hands, balance, posture, and contact with the ground.

The overlap explains why automakers see robotics as adjacent rather than unrelated. They already employ engineers who work across sensors, embedded computing, safety systems, electric motors, batteries, simulation, and large-scale software deployment.

GAC offers a direct example. Its GoMate system uses a vision-based mapping approach for navigating and understanding physical space, according to the company’s robot program.

The company also says GoMate can switch between remote and autonomous control. That hybrid model is common during early deployments because human operators can handle situations beyond the machine’s current abilities.

Remote intervention is not evidence of failure by itself. It becomes a problem when intervention is frequent, expensive, or hidden from customers.

Automakers also understand energy management. Battery size, thermal behavior, motor efficiency, power electronics, and charging strategy shape both electric vehicles and mobile robots.

Weight creates another shared constraint. Adding a sensor or computing module affects energy consumption, heat, cost, and mechanical design. Automotive engineers already make those tradeoffs at product scale.

Supply chains may be an even larger advantage. A humanoid requires actuators, reducers, bearings, sensors, wiring, battery systems, processors, structural parts, and safety hardware.

Car companies know how to qualify suppliers, control component variation, and manage thousands of interacting parts. They can also redesign components for automated assembly.

This experience matters because an impressive prototype is not the same as a repeatable product. Small manufacturing differences can change joint friction, calibration, balance, heat dissipation, and battery performance.

Automotive quality systems are designed to detect such variation. They track parts across production, document faults, and connect field failures to specific manufacturing batches.

Software operations also transfer. Connected vehicles receive updates, generate diagnostic data, and support remote maintenance. Deployed robots need similar systems for fleet monitoring, incident review, and software rollback.

Yet a humanoid is not a car with arms. Vehicles generally move on prepared roads governed by traffic conventions. A legged robot encounters deformable objects, uncertain contact, stairs, clutter, and people who behave unpredictably.

Manipulation is especially difficult. A robotic hand must manage shape, texture, weight, force, and slippage. A minor error can damage an object or create a safety risk.

This is where embodied AI departs from familiar vehicle intelligence. The machine’s actions change its observations continuously, and contact introduces uncertainty that cameras alone cannot resolve.

Automakers possess many useful building blocks. Their task is to prove those blocks can be recombined into reliable machines rather than attractive demonstrations.

Manufacturing Scale Meets Robotics Specialization

Automakers can accelerate robot production, while robotics specialists retain deeper experience with bodies, control systems, and real-world instability.

Unitree remained one of the conference’s most visible robotics companies. Its machines boxed, danced, and played table tennis, offering a direct display of balance and coordinated motion.

Such performances can appear frivolous. They still stress capabilities that matter elsewhere, including fast recovery, timing, whole-body control, and stability under changing loads.

Robotics specialists have spent years working on these problems. Their engineering teams often focus narrowly on locomotion, joint design, actuator control, reinforcement learning, and teleoperation.

Automakers bring a different concentration of expertise. They know high-volume manufacturing, product validation, supplier management, and complex safety processes.

The contrast creates the article’s main tension. Robotics companies know how to make machines move, while automakers know how to turn difficult machines into industrial products.

Neither capability is sufficient alone. A stable prototype without affordable production remains a limited platform. A well-produced machine that requires constant intervention remains an expensive appliance.

Chery is trying to bridge that gap through scenario specialization. Aimoga is developing different physical configurations for traffic, medical, retail, companionship, and guided-service roles.

That approach limits the operating environment. A wheeled hospital guide does not need the balance capabilities of a biped crossing uneven outdoor terrain.

It also narrows training requirements. The robot can learn a defined route, a fixed service catalog, and a predictable set of interactions.

This is less dramatic than promising a general household humanoid. It is also closer to how successful automation usually enters a market.

Industrial robots became valuable because factories reorganized tasks around them. Floors, fixtures, components, and safety barriers reduced the uncertainty a machine had to manage.

Service robots will probably follow a related path. Early deployments can succeed when companies modify spaces, workflows, and employee responsibilities around the robot.

Dealerships are an obvious starting point for automakers. They contain branded information, controlled layouts, staff support, and repetitive customer questions.

BYD’s first humanoid was designed for that type of environment. The company said the machine would interact with visitors rather than remain a concept exhibit, according to its Di Space plan.

The wording still leaves important questions. Public interaction does not reveal how long the robot operates autonomously, how often staff intervene, or what happens when visitors behave unexpectedly.

Factories provide a different test. They have structured tasks, measurable output, trained employees, and established safety controls.

They also expose robots to repetitive stress. A machine that completes a short demonstration might fail after hours of continuous operation or thousands of repeated movements.

Joyson Electronics illustrates how automotive suppliers are moving into this layer. The company has worked on dexterous hands, electronic skin, robot head assemblies, controllers, and industrial training.

Its Ningbo automotive safety plant has hosted humanoid systems for industrial adaptation and data collection. Joyson describes the work as a route toward robot components, secondary development, and factory deployment.

The company’s industrial training highlights another automotive advantage. Suppliers can test robots inside operating plants while developing components for robot manufacturers.

This creates several possible roles. An automotive company can build complete robots, supply subsystems, provide training environments, or become an operator of robot fleets.

The winners might not be the companies with the most humanlike designs. They may be those that choose a narrow role and measure it honestly.

Automakers must also decide whether to compete with robotics specialists or supply them. Building an entire humanoid requires different economics from selling components across many platforms.

Partnerships can combine automotive scale with specialist control software. They can also create dependency, especially when core models, actuators, or operational data belong to different companies.

The competition will therefore unfold across multiple layers. Robot bodies attract attention, but components, deployment software, maintenance, and training data may determine commercial value.

The Demonstration Gap Remains the Hardest Problem

A robot becomes useful only when it performs safely for long periods under ordinary conditions, not when it completes a prepared routine once.

Conference floors favor visible success. Exhibitors can simplify tasks, reset objects, control lighting, restrict routes, and station engineers within reach.

Commercial environments remove those protections. Floors become wet, furniture moves, network connections fail, and people interrupt the machine.

A service robot also encounters linguistic ambiguity. Visitors change topics, provide incomplete instructions, speak over one another, or ask for services outside the approved workflow.

Large language models can make conversations more flexible. They do not eliminate mechanical limits, safety requirements, or uncertainty about the physical environment.

The central metric should therefore be sustained autonomous work. Buyers need to know how long a robot operates between human interventions and what those interventions cost.

A machine might complete most tasks while still demanding frequent remote assistance. That model can work, but the operator must count the humans behind it.

Reliability has several layers. The robot must remain mechanically functional, preserve balance, interpret commands correctly, and recover from unexpected events.

Safety adds another standard. A customer-service robot must avoid collisions. A factory robot working near employees must control force and enter a safe state after faults.

Humanoid form can increase the challenge. Two legs provide access to human spaces, but they introduce balance risks and greater control complexity.

Wheels often offer a more practical choice for flat indoor environments. They consume less energy and reduce the number of failure modes.

Aimoga’s mix of wheeled, humanoid, and quadruped systems suggests a scenario-first approach. The company is matching bodies to tasks instead of insisting that one humanoid perform everything.

That is a sensible engineering choice. It also weakens the popular idea that a general humanoid body will soon replace specialized machines across every setting.

Economics remains another uncertainty. A robot must create enough value to cover hardware, software, supervision, maintenance, training, and downtime.

Automakers may reduce hardware costs through scale. They cannot manufacture their way out of weak task economics.

A showroom robot that attracts visitors has marketing value. Measuring that value is difficult because attention can fade once the machine stops being novel.

Traffic and medical guidance offer clearer operational outcomes. Operators can track completed interactions, response times, route coverage, incidents, and human labor saved.

Even those deployments require careful comparison. A fixed kiosk, mobile screen, or conventional wheeled robot may perform the task more cheaply than a humanoid.

Companies should therefore compare robots against the simplest effective alternative. The relevant question is not whether a humanoid can perform a task.

The question is whether its mobility, manipulation, and social presence justify the additional cost and complexity.

Data claims also need scrutiny. A deployment can generate many hours of video without producing enough varied examples for meaningful model improvement.

Useful training data must capture decisions, actions, outcomes, errors, recoveries, and environmental context. It also needs consistent labeling and appropriate privacy controls.

Automakers have access to facilities and fleets, but access alone does not create high-quality embodied data. They must design collection systems around learning objectives.

The same caution applies to claims about autonomous operation. Companies should distinguish scripted sequences, supervised autonomy, remote operation, and fully independent execution.

Those categories can look similar in a promotional video. They imply very different labor costs and technical maturity.

None of these concerns means automakers will fail. They explain why the transition from car intelligence to robotics will take more than transferring a model or production line.

The conference established credible intent. Continuous field performance will determine whether that intent becomes a durable business.

What Comes After the Robot Showroom

The next phase will be decided by deployment evidence, not another round of more polished humanoid introductions.

The first signal to watch is operational disclosure. Automakers should report fleet size, working hours, task completion, intervention frequency, downtime, and safety incidents.

These figures would make deployments comparable. They would also separate routine operations from pilots that depend on engineers standing nearby.

Chery’s geographically distributed scenarios make it a useful test case. If Aimoga expands those deployments while maintaining performance, manufacturing-backed robotics gains credibility.

If expansion stalls, the results would suggest that adapting robots to each new environment remains expensive. That would favor narrower systems and specialized integrators.

The second signal is factory adoption. Automotive plants offer measurable work, controlled processes, and direct access to maintenance teams.

Robots that perform inspection, material handling, sorting, or repetitive assembly can generate clear evidence. Buyers can compare cycle time, failure rates, and labor requirements.

Joyson’s industrial training work points toward this route. BYD, GAC, Chery, and other manufacturers can also test machines inside their own production systems.

Internal adoption deserves careful interpretation. A company might subsidize a pilot for strategic reasons even when the economics would not satisfy an external customer.

Independent orders would provide stronger evidence. Repeat purchases would matter more than memoranda, demonstrations, or small evaluation batches.

The third signal is product architecture. Automakers must reveal whether they are building general humanoids, scenario-specific fleets, component businesses, or integrated service contracts.

Each model requires different capabilities. A component supplier needs compatibility and production quality. A fleet operator needs maintenance, monitoring, and customer support.

A general humanoid program carries the broadest promise and the highest risk. It must combine mobility, manipulation, reasoning, safety, and acceptable cost within one platform.

Scenario-specific systems can reach useful work sooner. However, they may produce fragmented product lines that lose some manufacturing efficiency.

Component strategies can spread risk across several robot makers. They also place the most visible customer relationship elsewhere.

The fourth signal is partnership structure. Automakers possess factories and distribution networks, while robotics specialists hold valuable control systems and deployment experience.

Deep partnerships could compress development cycles. Weak partnerships might leave companies dependent on external technology for the most important parts of the product.

Ownership of operational data will be especially important. The organization controlling field data can improve models, identify failures, and shape future hardware.

Finally, watch whether robots expand beyond locations controlled by their parent companies. A dealership or corporate campus is useful, but it does not prove demand from outside buyers.

Hospitals, municipal agencies, logistics operators, retailers, and independent factories will impose harder procurement standards. They will compare machines against existing automation and human-operated services.

Those customers also demand maintenance commitments. Robot vendors need spare parts, trained technicians, remote diagnostics, cybersecurity processes, and reliable update systems.

Automakers already operate large service networks. Repurposing those networks for robots could become one of their strongest advantages.

The transition will not be automatic. Vehicle technicians require new training, while robot failures involve unfamiliar combinations of software, sensors, joints, and human interaction.

China’s automakers have now made their direction visible. They are moving beyond financing robotics startups and treating embodied systems as products, components, and operating platforms.

That strategy places pressure on specialist robot manufacturers. It also forces car companies to compete outside the carefully standardized world of vehicle production.

For developers, buyers, and AI teams, the useful response is disciplined observation. Track autonomous working hours, intervention rates, external orders, and repeat deployments.

Do not treat a humanlike body as evidence of general intelligence. Do not dismiss automakers merely because their first jobs involve greeting visitors or guiding customers.

Those bounded assignments are where companies can collect operational data and discover which engineering advantages really transfer. They are also where weak economics become difficult to hide.

The question after Beijing is no longer whether car companies want to build robots. It is whether their manufacturing discipline can turn embodied AI into dependable work, one measurable deployment at a time.