This year, a batch of benchmark deployments for humanoid robots has been publicly revealed:
Three Figure 03 units are conducting live broadcasts of day-and-night uninterrupted parcel sorting in a logistics warehouse in the United States; eight Agibot Genie G2 units are conducting live broadcasts of parallel operations for 10 hours a day over 6 consecutive days on the tablet quality inspection production line at Longcheer Technology in Nanchang; Galbot S1 has achieved 7×24h continuous operations on the power battery production line at CATL...
Compared to the past two years, the most easily noticed metric in the humanoid robot industry has been "how many units were sold." However, if the lens is truly focused on factories, warehouses, retail stores, or other actual operation sites, it becomes clear that there is still a long way to go between purchasing a humanoid robot and truly acquiring a stable "digital employee" capable of working.
Data Source: MIR DATABANK
Therefore, an important shift is occurring in the industrialization of humanoid robots: the focus of industrial competition is gradually shifting from "who can build robots" to "who can turn robots into replicable productivity."
The First Step in Deployment Is Not Training Humanoid Robots, but Defining What Humanoid Robots Can Do
The first issue after humanoid robots enter real-world scenarios is: How can a human job be broken down into tasks that humanoid robots can understand, execute, and verify?
In the past, traditional industrial robots typically corresponded to clear processes with relatively stable task boundaries. Engineers could solidify movements through process parameters such as programs, trajectories, and takt times. However, the value of humanoid robots lies in their aspiration to enter more unstructured work that originally relied on humans.
For example, "warehouse order picking" can be broken down into at least the following:
Identify shelves and goods → Understand product categories → Determine placement locations → Plan walking and operation paths → Control arms and fingers → Grasp → Transport → Determine destination → Handle occlusions and collisions → Place → Inspect results → Handle exceptions
Moreover, exceptions in different scenarios vary completely. For instance, cardboard boxes may be deformed, goods may be occluded, shelves may have empty slots, target objects may be pressed by other items, and placement locations may change...
Therefore, it is crucial to define tasks well first for the deployment of humanoid robots. Specific definitions include: tasks, sub-tasks, reusable motion skills (as follows), exceptions, successful outcomes, and situations that need to be handed over to humans. This is essentially transforming the "job experience" previously held by humans into machine capabilities that robots can learn, invoke, and verify.
Listing Main Motion Skills of Humanoid Robots from a Physical Execution Perspective
Information Source: Compiled and collected by MIR Research & Industry
Humanoid Robots Currently Lack "Mass Production Capability for Real-World Data"
Once tasks are defined, what humanoid robots ultimately need to learn is how humans perceive, judge, move, operate, and achieve results in real environments. This means that during the deployment of humanoid robots, there is a continuous need to acquire real-world data. More importantly, this data production is showing a trend toward specialization and scaling.
On September 23, 2026, Mifeng Technology officially launched the "Mifeng Pai" physical AI (Artificial Intelligence) data crowdsourcing platform. The platform organizes users via an App to wear collection devices, complete designated tasks in various real environments such as retail, warehousing, manufacturing, and households, and convert real operations into robot training data. Public information shows that it covers 22 major categories of scenarios, over 5,000 tasks, and more than 50,000 real environments.
Workers at the Pearl Global garment factory in India wear Egolab AI custom camera rings for about 6 hours a day to record first-person hand sewing operation video data.
From this specialized, socialized, and platform-based data production method, it can be seen that the demand for real-world data in training complex motion models is continuously expanding. In the future, the collection and use of data will see a division of labor:
Scenario providers supply real environments;
Data platforms are responsible for organizing collection personnel;
Collection devices are responsible for recording human movements and environmental information;
Data engineering companies are responsible for cleaning, slicing, trajectory reconstruction, and quality control;
Model companies are responsible for training;
Ultimately, robot manufacturers or industry customers conduct post-training and deployment;
Deployed scenarios also become data sources for the continuous iteration of robot capabilities;
...
General Foundation Models Solve "Whether They Can", While Industry Application Models Solve "Whether They Can Do This Specific Job"
Once humanoid robots have data, the next step is model training. However, two stages need to be distinguished here: pre-training and post-training.
Pre-training addresses the "general capabilities" of humanoid robots. Through massive real-robot data, human demonstrations, teleoperation data, and simulation data, humanoid robot models aim to gain a basic understanding of environments, objects, actions, and tasks, and adapt to different humanoid robots, different objects, and different scenarios as much as possible. This is also the core of the current competition in humanoid robot foundation models.
However, the issue is that even if a general model has strong generalization, similar to humans' basic life and common task-handling capabilities, it may still not be immediately competent for various industry jobs. It still needs to be re-adapted for specific job SOPs (Standard Operating Procedures), onboarding training, and rule assessments. For example, a general model might know how to grasp a part, but what the factory truly needs is: identifying the part model, determining the orientation, completing the grasp, aligning with the assembly position, inserting it in place, and readjusting when deviations occur.
Therefore, after the foundation model, it is necessary to enter the post-training stage.
The core of post-training is to utilize massive real data from specific industries and specific tasks to further transform general capabilities into executable industry capabilities, which include task models, rule models, skill models, etc.
It can be seen that the key to future competition in industry deployment may not just be the model itself. Whoever possesses more real industry data, understands specific tasks better, accumulates more exception cases, and has more actual deployment experience, will more easily turn general models into real industry productivity.
Why Can't Humanoid Robots Currently Go to Work at Scale Directly?
Customer sites will never be prepared exactly according to the training-designed environments. The materials, lighting, workstations, takt times, safety requirements, equipment interfaces, and even the customers' definitions of "completing a task" at the site will all differ.
Therefore, a new industrial role is becoming increasingly important: humanoid robot secondary development and scenario deployment service providers. What they truly do is transform the robot's general capabilities into business capabilities that customers can use, specifically including sensor installation, end-effector adaptation, model post-training, model integration, motion orchestration, host computer development, and connection with the customer's original equipment and software systems.
Currently, humanoid robot body manufacturers need a secondary development ecosystem. The common types of secondary development manufacturers include:
Common Types of Secondary Development Manufacturers for Humanoid Robots
Information Source: Compiled and collected by MIR Research & Industry
Embodied intelligence algorithm model companies corely output software capabilities such as cross-body perception, task planning, and large model interaction. Body manufacturers master underlying motion control and body hardware resources, while algorithm model companies undertake the development of upper-level intelligent capabilities. Theoretically, the two form a natural division of labor, but the ecological game focuses on the "boundary of secondary development permissions" (such as ownership of scenario data, ownership of skill assets obtained through training, etc.). The division of these rights and responsibilities directly determines the distribution of benefits among all parties.
Peripheral hardware supporting companies use external hardware such as dexterous hands, mobile chassis, and machine vision as carriers. Relying on the open interfaces of humanoid bodies, they carry out secondary development, develop peripheral linkage logic and perception task programs, and achieve collaborative operations between the body and peripherals. These manufacturers simultaneously master component design, underlying driver adaptation, and perception algorithm capabilities. Their core is not just hardware supply, but completing the "body-peripheral" operation closed-loop through secondary development.
The core advantages of professional secondary development manufacturers for vertical scenarios are industry customer resources, on-site deployment experience, and process understanding. Most of these manufacturers currently come from the ecosystem partners of the original body manufacturers, and there is generally a binding relationship of interests. They do not necessarily develop embodied large models themselves, but they are familiar with the entire process of on-site engineering deployment. Their core purpose is to embed humanoid robots into existing business systems and deliver deployable scenario solutions for industries such as industrial, education, inspection, and entertainment. The challenge for these manufacturers lies in integrating humanoid robot capabilities with existing industry workflows and automation systems to achieve stable and scaled deployment.
Currently, MIR Research & Industry statistics show that the number of humanoid robot secondary development manufacturers exceeds one hundred.
Having Projects Does Not Equal Having a Market: What Is the Real Problem with Current Secondary Development?
Having many projects does not mean the market is already mature. Currently, clients mostly purchase the display value of humanoid robots, while rarely purchasing the hard-to-achieve productivity value. Therefore, different downstream scenarios show different levels of commercial maturity in secondary development.
Overview of the Current Status of Secondary Development Deployment in Different Downstream Scenarios
Information Source: Compiled and collected by MIR Research & Industry
Projects in these three types of scenarios may all generate revenue, but the purposes of deployment differ.
In entertainment performance and reception/guiding scenarios, the humanoid robot itself is the content of the display demand. Such projects can generate revenue and help clients build market influence. In industrial and complex commercial service scenarios, clients mostly want to verify whether humanoid robots can perform tasks. Such projects can be deployed but may not be quickly replicated. C-end (Consumer-end) operation projects are currently difficult to open up in the market, with most weighing the trade-offs between functions and prices.
The commercialization of secondary development in industrial, complex commercial service, and C-end operation scenarios is difficult, mainly because their process/task complexity, safety costs, and replication costs are all high. Among them, "how to make the second project cheaper than the first" is the focus of market promotion. A truly mature secondary development market needs to start seeing reusable assets: data accumulated from projects, models trained from projects, skills developed from projects, interfaces established from projects, safety solutions completed from projects, etc.
It is only when secondary development manufacturers start selling a skill package or a set of industry solutions that the secondary development market truly begins to have business scale effects.
Conclusion: What Is Truly Worth Observing in the Next Stage of the Industry Is the "Scaled Deployment Capability of Robot Scenarios"
In 2026, China's humanoid robot industry continues to advance the process of scenario deployment and is undergoing a difficult transition: transforming from a hardware product into a productivity system that can be continuously trained, developed, deployed, and operated.
The humanoid robot industry is currently answering: "Can robots actually work?" Moving forward, the industry must answer: "Can this working capability be replicated at low cost to more factories, more positions, and more real environments?" This requires body manufacturers to connect the chain of "components, production, data, models, development and deployment, services, and standards," maintaining multi-party collaboration in the industrial chain. The maturity of this system determines whether body manufacturers can enter the real productivity market.