A component factory with a second purpose

On October 20, 2026, Schaeffler opened a roughly 7,500-square-meter humanoid robotics factory in Taicang, Jiangsu, China. Six humanoids work across five stations, preparing and sorting parts, joining and inspecting them, then labeling and packaging them.

The products are themselves key components for humanoids, including actuators that drive their joints. In that specific sense, robots are assembling parts for the next generation of robots. Schaeffler describes the facility as the start of mass production of these components and has reported approximately €400 million in global humanoid orders.

But the site is not simply more manufacturing capacity. It also includes a training and data-acquisition center, a digital-twin platform and a data center. Why would a component supplier put robot training alongside production?

Two subsequent announcements sharpen the question. On October 21 Schaeffler announced actuator simulation work with NVIDIA. On October 22 it announced a Wandelbots partnership aimed at developing robotic tasks that could move between production sites.

Building robots, teaching them useful factory work and making that work reusable elsewhere are distinct challenges. Schaeffler is extending its activities across those boundaries.

1. What 'humanoids making humanoids' actually means

Schaeffler has not announced a factory that autonomously assembles complete humanoids from beginning to end. The stated output is key joint modules, including linear and rotary actuators. These convert electrical energy into controlled mechanical movement and help coordinate a robot's limbs and handling tasks.

The company is applying precision-motion and volume-manufacturing expertise from automotive and industrial markets to robot components. On the Taicang line, the six humanoids participate in staging, sorting, joining, inspection, labeling and packaging at five stations.

A digital twin mirrors the actual line. Sensors collect assembly accuracy, task duration, energy consumption and equipment-condition data. Those measurements can support evaluation, but the announcement does not publish line throughput, defect rates, actual utilization, intervention frequency or per-robot productivity.

It therefore does not establish how many people the robots replace. Nor does it establish the extent of remote operation or the autonomy of each task. Real participation in component production is confirmed; the economic performance of that production line is not.

The important structural choice is to put production and training in the same location, rather than to treat a visible robot demonstration as proof of commercial returns.

2. €400 million in orders does not remove the need for training

In its October 13 preliminary results announcement, Schaeffler said humanoid orders had risen to roughly €400 million, helped by two new customers in Greater China. Its October 20 factory announcement repeated that figure.

This is the reported order scale of the global humanoid business, not revenue at the Taicang factory. It must also be distinguished from recognized sales, operating profit and cash collected. It signals customer demand for components, not the efficiency of the assembly line.

Factory work introduces another problem. To pick up a part, a robot must locate it, determine its orientation, apply suitable force and place it correctly. It must handle variation in the initial position while maintaining quality and equipment safety through repeated cycles.

Hands, joints, sensors and control software must work together under those conditions. Buying a robot does not automatically deliver that capability. The user must implement and validate work for its own environment.

Taicang is designed to manufacture components while testing this application problem. The training center addresses a different constraint from component-production capacity.

3. The training unit may be a reusable task, not one robot

Schaeffler identifies visual inspection, packaging, material handling, container transport and replenishment among the tasks covered by its training and data-acquisition center. These may attract less attention than a striking demonstration, but they recur in manufacturing.

Container movement, sorting, packaging and feeding materials to a line can be relevant at multiple plants. Schaeffler says the training environments reflect conditions across approximately 100 production locations.

Teaching one robot one task requires a suitable training environment. Reusing comparable work across factories requires more: retaining the method in data and software, adapting it to a new setting and validating it again.

The company plans to connect training, simulation and production data to improve robotic performance. Banseog interprets this as an effort to make accumulated task experience reusable, rather than only to improve an isolated machine.

That interpretation is not evidence that every task has already been deployed at 100 factories. Having a training facility and reproducing the same operating result across a production network remain different achievements.

4. Schaeffler is both a component supplier and a robot user

Schaeffler did not begin as a finished-humanoid startup. Its background in bearings, precision engineering and motion technology supports an actuator business. It is also seeking to use the robots made by its partners in its own factories.

A January 2026 partnership with the UK company Humanoid combined actuator development and supply with a plan to introduce several hundred humanoids across Schaeffler's global production sites over five years. An April partnership with Hexagon Robotics envisaged at least 1,000 humanoids over seven years.

These separate plans should not be added together and presented as confirmed purchases or installed robots. Orders, deliveries and operational deployments would require separate evidence.

Nevertheless, the dual role matters. A component supplier needs to understand the conditions in which its actuators operate; a factory customer needs reliable task performance. Experiencing both could connect application feedback with product development.

For example, a joint-temperature or force-control problem observed during a task could inform actuator design and validation. That is a possible learning mechanism, not a measured improvement disclosed by Schaeffler. Supplying a component and also experiencing its industrial use can provide different information.

5. NVIDIA: test joint physics before the robot reaches hardware

On October 21, the day after the factory announcement, Schaeffler announced work with NVIDIA on high-fidelity actuator digital twins. The collaboration involves the Newton physics engine and NVIDIA Isaac Lab and Isaac Sim.

The question is how faithfully simulated movement represents real hardware. A virtual robot may lift a particular object, but an actual joint has torque, speed, friction, thermal and mechanical constraints. An incomplete model can produce behavior that succeeds in simulation and fails on the machine: the sim-to-real gap.

Schaeffler intends to model its actuators closely enough for developers to examine behavior before installing the physical components. A successful implementation could reduce repeated hardware experiments and some validation effort when related actuators are used in another robot model.

One long-term objective is zero-shot deployment, reducing extensive additional training in the physical environment. It is a technical ambition, not evidence of a universal, commercially achieved capability.

Even an accurate actuator model does not automatically validate factory lighting, object positions, surrounding equipment or safety conditions. Better joint simulation addresses one boundary; reusable industrial tasks address another.

6. Wandelbots: transfer the task, not just the joint model

On October 22, Schaeffler announced a technology partnership with Wandelbots around the NOVA platform. It supports application development and control across robot systems in a common software environment.

The partners aim to develop tasks, test them digitally and deploy them in production. More importantly, they want work validated at one site to be usable at other factories and on other robotic systems. A proposed skill store would collect applications, models and technical knowledge.

The three roles should remain distinct. NVIDIA Newton and Isaac address representations of physical joint movement and constraints. Wandelbots NOVA connects application logic, models and simulation for transfer between systems and locations. Taicang provides real manufacturing conditions for testing tasks, component quality and data collection.

These announcements do not establish that all three elements are already a fully integrated commercial system. They show complementary efforts within the same corporate strategy.

A physically feasible motion is not necessarily a task that can be repeated reliably in a factory. Schaeffler is pursuing ways to narrow both gaps, rather than treating them as the same engineering problem.

7. Why storing a task can matter economically

Production environments differ. Workbench height, part geometry, grippers, tools, neighboring machines and separation from people can all change. A task that already works at one plant may need new configuration and testing at another.

When that work repeats across plants, the customer repeatedly pays not just for robots but for application development and validation. That cost is separate from the purchase price of the hardware.

Consider a sorting task. Another factory may need similar sorting, yet use a different robot or part-feeding arrangement. Preserving core application logic and validation knowledge could let engineers adapt only what changes instead of solving the whole problem again.

This is the proposed value of reusable robotic skills. A stored skill, however, is not automatic compatibility with every robot and plant. Reach, tooling, environment, controllers and safety conditions must still be assessed.

The useful measure is therefore not the size of the code or data library. It is how much development and verification effort is genuinely avoided at the next deployment.

8. An existing packaging case reports about two weeks saved

Wandelbots describes an earlier robot-assisted packaging application in which simulation-based development cut development time by around two weeks per application. Processes, cycle times and quality inspections were checked in a digital twin before physical deployment, reducing later correction work.

This is a supplier-reported result from a particular earlier project. It is not a productivity result from the new Taicang humanoid line, and it is not proof that every automation project saves two weeks.

It must also be kept separate from the partners' goal of reducing automation life-cycle costs by up to 40%. One is a reported development-time example; the other is a future cost target.

The disclosed materials do not independently establish the financial savings or generalizability across industries. Nevertheless, the mechanism is relevant: shortening the work needed to put a robot into productive use can create value without lowering its purchase price.

That is a different form of productivity improvement from increasing the number of components a supplier can manufacture.

9. How should the 40% cost target be tested?

Schaeffler and Wandelbots target up to a 40% reduction in automation life-cycle costs, covering both new and existing production environments. The announcement does not identify a plant that has already achieved that reduction or provide the baseline, comparison period, investment accounting and verification method.

Banseog proposes separating six questions. Development time: how long do the first and subsequent applications take? Cross-site transfer cost: does adapting the same task at another plant require less work? Task success: are quality and repeatability preserved under real conditions?

Human intervention: how much effort is needed to correct errors or restart work? Utilization: does shorter development coincide with more operating interruptions? Total life-cycle cost: do hardware, software, integration, training and maintenance together become more economical?

These are Banseog's analytical criteria, not official KPIs published by Schaeffler. They distinguish making a task work at the first factory from making it work more cheaply at the second.

A global manufacturer must evaluate the repeated cost of training, configuration and validation as it adds locations. Reducing that repeated burden could turn isolated automation projects into a more repeatable production system.

10. Application experience may strengthen the component business

A supplier that installs, trains and transfers robots in its own factories could learn not just how a joint performs technically, but what the complete robot must deliver for an industrial customer.

Robot manufacturers face heating, repeated loads, control response and long-term maintenance as well as headline actuator performance. If Schaeffler can feed actual operation data into component validation and development, it could offer more concrete design information and reliability evidence.

Simply improving a component specification without understanding the factory task may not solve the customer's deployment problem. The strategy can therefore be read as building two assets: hardware for precise motion and experience validating useful work with that hardware.

The announcements envisage spreading internally developed skills to other factories and potentially making them available externally. They do not establish that an external skill-library business is already selling products or generating software subscription revenue.

The open business question is whether a component manufacturer can also create value by making robots easier to adopt. That possibility should not be mistaken for a completed commercial model.

11. Does the work have to be done by a humanoid?

The fact that humanoids participate in Taicang's line does not mean that sorting, assembly and packaging are exclusive to humanoid robots. Fixed industrial robots, collaborative robots and dedicated automation can perform some of those tasks.

A demonstration cannot establish that a humanoid is cheaper or more reliable than those alternatives. Its potential advantage may be versatility in workplaces originally organized around people. Whether that versatility produces savings is a separate question.

For a long-running, identical assembly task at a fixed location, dedicated equipment may be more stable or less expensive. Where workstations and task types change frequently, a reconfigurable robot may have advantages.

The choice depends on production volume, changeover frequency, product quality and safety requirements. The test is not merely whether a humanoid can perform work, but where it is the more rational automation choice.

The training and data facilities could support such comparisons. Follow-up operating evidence is still needed to demonstrate economics acceptable to industrial customers.

12. The capability requirements extend beyond a robot engineer

Joint and drive development involves precision mechanical design, actuators, power electronics and motor control. Training robots in factories adds task-data acquisition, behavior learning, simulation, systems integration and production-quality understanding.

Moving a task between plants can also require application engineering, robot-interface adaptation, safety validation and field deployment. These capabilities are related, but not interchangeable.

An important role is understanding which operations recur, under what conditions they fail and how to translate them into work a robot can perform. Developing an AI model alone does not resolve those questions.

The announcements do not prove that Schaeffler is hiring for each of these roles or has substantially expanded its robotics workforce. New vacancies and organizational plans would be needed to establish actual labor demand.

The evidence supports a direction for the technical and operating system being built. Its possible capability implications are analysis, not confirmed recruitment activity.

13. What remains unproven

Several material questions remain. Actual actuator output, yield and cost at Taicang have not been disclosed. Neither has the extent to which the six humanoids operate without supervision or remote assistance. The conversion of roughly €400 million in global orders into deliveries and revenue requires further evidence.

There is not yet sufficient operating evidence showing how much actuator twins reduce customers' development and validation costs, or how consistently NOVA-based tasks transfer across different robots and environments.

The 40% life-cycle cost reduction remains a goal, not a result achieved throughout the production network. External skill-store prices, contracts and revenue models are also not established.

It would therefore be inaccurate to claim that Schaeffler has completed large-scale autonomous humanoid operation or can immediately send a task learned at one factory to every factory worldwide.

The narrower, verified change is a component-production facility combining training, data and validation, alongside actuator-physics simulation and cross-site application-development partnerships. Their contribution to costs and commercial performance remains to be tested.

BANSEOG VIEW

BANSEOG VIEW | The next contest does not end with how many humanoids were built

Robot price, shipments, joint specifications, task success and hourly cost relative to human labor all matter. Schaeffler's strategy adds another question: how much must be spent again to apply one robot's successful work to another robot and another factory?

If each location requires near-total redevelopment, deployment costs recur as the network grows. If part of a validated task can be reused and adapted, the same experience may become useful across multiple sites. Hardware prices alone do not reveal that difference.

Banseog interprets the combination of Taicang production and training, NVIDIA actuator simulation and Wandelbots application reuse as an attempt to reduce that repeated effort. Two possible sources of value follow: strengthening components through application-based reliability evidence, and improving the economics of robot adoption itself.

Neither possibility proves a new software revenue model. The evidence to watch is the actual time and money saved on subsequent deployments, preservation of task quality and safety across robot models, and expansion from pilots to sustainable production.

A humanoid's industrial value is not settled by moving successfully once. It becomes more significant when that experience can be repeated economically elsewhere. The Taicang factory is a place to test that possibility.

The most important element is therefore not simply six moving robots. It is the attempt to make their validated task experience useful for the next robot and the next factory.

SOURCES

Primary sources and references

As of October 26, 2026. Approximately €400m refers to reported global humanoid orders, not Taicang revenue. Six robots participating in work does not establish full autonomy, yield, productivity or return on investment. The two-week saving is an earlier vendor-reported packaging-application example; up to 40% life-cycle savings is a future target. Zero-shot deployment, cross-site economics and external skill commercialization remain prospective. Capability implications are analysis, not verified hiring.