The Next Battle in Embodied Intelligence Is Not in the Robot Itself

Gasgoo Munich- In a recent showcase at MHERO's general assembly plant, a domestic heavy-duty robot hoisted a 300-plus-kilogram M817 frame, handling its automated transfer and line loading with ease.

Within that same facility, the paint shop employs automated inspection and grinding robots, while humanoid robots are already in routine operation on the assembly SPS logistics line. AGVs manage transport, heavy-duty robots shift frames, specialized machines handle precision tasks, and humanoid robots bridge the flexible gaps.

This reality reflects the state of modern manufacturing far more accurately than the endless debate over whether factories actually need humanoid robots.

Factories are already selecting robots based on specific tasks. The real challenge arises as these machines proliferate: how to integrate them into a single production ecosystem.

On September 9, at the Convergent Intelligence Industry Development Conference (2026), Jin Bing, former first-class inspector and deputy director of the Policy and Regulations Department of the State Post Bureau, called for integrating unmanned equipment across the board—robots, autonomous vehicles, drones, and unmanned vessels. He outlined a vision of "one brain, one base."

Here, the concept of "one brain" refers closer to task-level coordination and scheduling, rather than forcing every robot to run on a single universal model.

As robots truly begin to enter factories, the competition is no longer confined to the hardware itself.

Robots Know How to Work; Factories Are Worrying About How to Use Them

The scenarios disclosed by MHERO this time illustrate the point well.

The heavy-duty robots boast a 750-kilogram rated payload and a 3.6-meter reach, with positional repeatability held within ±0.15 millimeters. They handle the cross-process transfer of frames weighing hundreds of kilograms. Meanwhile, the paint shop's automated inspection and grinding robots focus on identifying and locating paint defects, then polishing them. The humanoid robots, stationed on the assembly SPS logistics line, tackle more flexible grasping and transport tasks.

Different robots are judged by entirely different metrics.

Heavy-duty robots are measured by load, precision, and uptime. Grinding robots are rated on defect detection rates, smoothing efficiency, and polishing precision. For humanoid robots, the key metrics are grasp success rates, cycle times, and the ability to switch between workstations.

MHERO's data reveals that its humanoid robots' grasp success rate has climbed from 90% to 95.6%, while the cycle time per task has been slashed from 65 seconds to 20.

On a real production line, the so-called "advanced nature" of a robot quickly boils down to simple numbers: can it lift the weight? Is it accurate? Can it run continuously? Does it keep up with the beat? And how fast can it recover from a glitch?

At trade shows, fast running, high jumping, and complex moves grab headlines. But in a factory, a robot is first and foremost a piece of production equipment.

If a robotic arm can handle a fixed station, there's no need to force a humanoid into the role. For flat-ground logistics, AGVs and AMRs are usually the more mature choice. For transferring 300-kilogram frames, humanoids aren't the optimal solution either. Their advantage only becomes clear when workstations vary, tasks aren't standardized, and the machine needs to navigate existing human environments and tools.

So, manufacturers today aren't really agonizing over "what a robot should look like", they care about who does the job most efficiently.

The problem is, once robots find their respective places, a new complexity follows.

Take MHERO's current project. It simultaneously built an industrial scenario database, accumulating visual samples, operational actions, and abnormal condition data. As the project team noted, the real value lies not just in getting several types of robots into the factory, but in integrating robots, AI algorithms, machine vision, industrial software, and production systems into a unified whole.

This step shifts the question from "can robots do the work?" to "how does the system organize it?"

Scheduling Thousands of Robots Doesn't Mean Different Types Can Collaborate

Robot scheduling is nothing new.

The warehousing industry already manages hundreds, even thousands, of AMRs simultaneously. Some top-tier systems boast the ability to schedule clusters of several thousand robots in a single facility.

But there is a subtle distinction that is easy to overlook.

Managing 5,000 robots with similar capabilities is a completely different problem from managing 20 that are totally distinct.

In the past, robot scheduling mainly involved solving for location, battery level, route, and task allocation.

Which unit is closest to the shelf? Which one needs charging? Which route is blocked? The backend just needs to calculate the math. Because fundamentally, these robots possess similar capabilities.

As embodied intelligence advances, factories may soon host AMRs, robotic arms, humanoid robots, quadrupeds, and various specialized machines—all from different vendors.

The system needs to know not just where a robot is, but exactly what it can do.

Take "grasping," for example. Some robots can only handle standard cardboard boxes; others manage soft bags; still others perform precision assembly. Similarly for "movement": AMRs are for flat floors, quadrupeds can climb stairs, and humanoids might cross obstacles or operate equipment.

When a production task comes in, it will likely first be broken down into transport, recognition, grasping, assembly, and inspection—then assigned out based on the specific capabilities of different robots.

If one link fails, the system must decide: retry in place, switch to a different robot, or reschedule the entire subsequent workflow.

This is no longer "scheduling vehicles" in the traditional sense.

Moving from "which machine is free" to "which machine is best suited to get this done" requires a completely different set of system capabilities.

PIA Automation's automotive BMS multi-robot collaborative pilot line, showcased this year, already demonstrates this kind of layering.

Six robots cover raw material transfer, precision assembly, labeling and packing, rack stacking, patrol inspection, and predictive maintenance, all under the global control of the line's PLC. Ecosystem partners provide the bodies, interfaces, and base models, while PIA Automation handles process reconstruction, skill encapsulation, line orchestration, full-system integration, and mass production validation.

However, this remains a pre-designed production line. The equipment, interfaces, and process flows have all been adapted for this specific project.

It proves one thing: the robot hardware and the "organizer of the robots" can already be decoupled.

But turning this separation from a custom line into a replicable system capability means solving a harder problem—can processes, skills, interfaces, and task descriptions be standardized and encapsulated so they can be reused at the next factory?

That is the true threshold for moving from "multi-machine collaboration" to "heterogeneous collaboration."

Everyone Wants to Be the "One Brain"

If this capability is truly realized, a new battleground emerges: who controls the task orchestration and scheduling?

Robot manufacturers naturally hope the answer is themselves.

The hardware is theirs, the model is theirs, the data is theirs—so they'd prefer to keep the scheduling system in-house too. The interfaces align best, performance is easiest to optimize, and customers buying the whole package simplifies integration.

Tesla is a direction worth watching.

Optimus is being integrated into Tesla's own manufacturing and AI ecosystem. The robot body, model, chips, and factory scenarios are all being advanced through internal collaboration.

The advantages of this vertical closed loop are tangible: no need to move data across companies, hardware and software can be tweaked together, and problems are easier to diagnose.

Still, this closed loop is in its early stages. Optimus is still ramping up production and undergoing internal application validation. It points to a possible direction more than a proven result.

More importantly, most factories aren't as "neat" as Tesla's.

Equipment in a car plant might come from dozens, even hundreds, of suppliers. The last generation of lines hasn't retired yet, and the next generation of robots is already arriving. Future humanoid robots won't all be bought from a single company either.

This leaves room for another path: horizontal openness.

Industrial software firms can descend from the MES and WMS layers to the equipment layer; system integrators can move up from the line; AI companies hold the models; cloud platforms control the dev environment; and large manufacturers might even build their own layer.

Everyone's motive is roughly the same: whoever controls this layer isn't just selling products—they are defining how other products are used.

But the biggest rival for a horizontal platform isn't another platform—it's the closed loop itself.

If leading robot firms and major manufacturers generally choose to lock their hardware, models, data, and scheduling inside their own ecosystems, the market space for third-party horizontal platforms will be significantly compressed.

For robot companies, being closed isn't necessarily bad.

The deeper the interface binding, the higher the cost for customers to switch. Keeping data in-house accelerates model iteration. Selling hardware, software, and services together often yields fatter margins.

Factories, however, think the opposite. Buyers generally don't want to be locked in by a single supplier—especially in manufacturing, where equipment sources are complex and lifecycles are long. They want new robots to plug into old systems and the freedom to choose other suppliers in the future.

So, the real competition in the next phase may not be "third-party platforms replacing robot makers", it is more likely to be a long-term tug-of-war between two ecosystems, one chasing closed-loop efficiency, the other pursuing openness and compatibility.

Who gains the upper hand depends on what factories care about most in the end: the extreme efficiency of a single system, or the long-term freedom to choose suppliers.

Standards Are Just the First Step; The Hard Part Is Who Is Willing to Open the Door

Horizontal collaboration has one unavoidable prerequisite: everyone must first be able to "speak the same language."

The "Industrial Mobile Robot Scheduling System Underlying Data Interface Requirements" (GB/T 47864-2026), released in July, has already begun regulating industrial mobile robot scheduling interfaces at the national standard level.

The "Technical Specification for Humanoid Robot Integration Multi-Machine Management Systems" has also entered the national standard planning stage, covering robot access, tasks, monitoring, collaboration, management, and operations.

Standards address more than just "how to connect interfaces." The real issue is whether devices from different vendors can be understood by the same system.

Only when equipment status, task definitions, skill descriptions, and exception mechanisms gradually form a common language can cross-brand scheduling escape the cycle of "re-adapting for every new project."

Taking another step, we might even see a so-called "robot capability marketplace."

A factory might not need a robot from a specific brand, but rather a specific capability like "move 100 kilograms of material," "perform millimeter-level assembly," or "conduct continuous inspection for 8 hours." Whoever can do it, the system calls.

At that point, robot companies would be selling not just hardware, but callable skills.

It sounds like a technical issue, but the real difficulty isn't entirely technical.

Why should robot makers open up their most core interfaces?

If standardizing interfaces makes it easier for clients to switch suppliers, top vendors with scale, brand recognition, and a customer base may lack the motivation to open the door wide.

Industrial software vendors, integrators, and platform companies, on the other hand, want interfaces to be as unified as possible.

The easier it is to connect equipment, the higher their platform value.

So, behind standards lies a redistribution of interests.

How open to be, who gets to see what data, who gets to call which capabilities—these won't be decided by technology alone.

Scheduling Rights Don't Natively Equal Data Rights

If a system truly connects many different robots, it theoretically sees a picture that no single robot vendor can see.

Which task was assigned to which robot?

Why did it fail the first time?

Why did it succeed after switching to another robot?

Which process frequently requires human intervention?

Which equipment is most efficient in what environment?

As this information accumulates, the platform knows not just "where the machine is," but gradually "which machine is best suited for what."

This is where the most exciting potential for the scheduling layer lies.

But we shouldn't assume this will go smoothly.

Scheduling platforms don't naturally own this data.

In real commercial projects, production data may belong to the factory. Robot makers might restrict secondary use of operational data. Data involving core processes might require local deployment, never leaving the customer's server.

Even if the platform handles scheduling, it doesn't have the right to take that data and train its own models.

So, for "scheduling rights to become data rights," a critical precondition must be met: data usage rights, anonymization mechanisms, and contractual boundaries must allow the platform to truly accumulate cross-hardware experience.

Without this, the platform might hold only real-time control, not a long-term data moat.

This complicates the competition in the next phase.

It's not just about models and interfaces; it's about who can convince clients that plugging more equipment into their system won't mean losing control or data sovereignty.

Today, what catches everyone's eye is still the robot hardware. Who runs more steadily? Whose dexterous hand is more agile? Who dropped the price below 100,000 yuan? These make the news.

But a factory ultimately doesn't need 100 of the "smartest robots", it needs these machines to finish tasks on time, on beat, and to spec.

These are not the same thing.

Individual machine intelligence determines if a robot can get the job. System capability determines if a swarm of robots can truly form a productive force.

The next battle in embodied intelligence will, of course, still be fought over hardware, dexterous hands, models, and core components.

But a new front has opened up.

Whoever can organize robots, tasks, interfaces, and data into a replicable system will have the chance to capture the greater share of value that comes after robots reach scale.