Key Takeaways:
- The article argues that the next industrial divide will be decided task by task, as countries race to deploy adaptable robots reliably and cheaply rather than awaiting general-purpose humanoids, citing DexForce’s simulation work and a reported Germany–Nvidia industrial-AI plan.
- It notes that 5.079 million industrial robots were operating worldwide in 2025 with 603,000 new installations, and that bounded pilots such as BMW’s Figure 02 handling over 90,000 parts show value in specific handoffs, not wholesale job replacement.
- The author concludes that effective policy should measure cost per accepted unit of work, link training to real jobs, and expand testing access for smaller firms, because employment shifts may appear first as unfilled vacancies or reorganized shifts rather than headline layoffs.
China is investing in physical AI, Germany is preparing new industrial partnerships, and factories are testing humanoids on narrow jobs. The next competitive advantage may arrive long before robots become general-purpose workers.
In Shenzhen last week, a Chinese robotics startup showed how a virtual robot can practice moving objects through many simulated variations before a physical machine attempts the task. In Berlin, government sources said Germany was preparing an agreement with Nvidia to expand industrial AI, including tools for training robots. The stories concern different technologies and business models. They point toward the same industrial contest: which countries can turn adaptable machines into dependable work at an acceptable cost?
Reuters reported the Chinese firm DexForce’s approach on October 15. The German initiative was still a reported plan on October 16, not a completed agreement. Neither event establishes that humanoid robots have become general-purpose employees. Together, they show why the strategic threshold is changing. A factory does not need a machine capable of every human task. It needs one that can perform a specific part of production reliably enough to reorganize the work around it.
The next industrial divide may therefore emerge one task at a time, before a spectacular breakthrough in general-purpose robotics and before the public sees a wave of mass layoffs. Industrial policy and workforce planning are still largely organized around factories, firms and occupations. Automation advances at the level of individual tasks.
The race is already bigger than humanoids
The scale of established automation is easy to mistake for evidence that humanoid workers have arrived. The International Federation of Robotics reported in September that 5.079 million industrial robots were operating worldwide in 2025, with 603,000 new installations during the year. Its separate professional-service-robot figures showed about 250,000 shipments in the sample, with transport and logistics the largest application. These are different categories and measures, and neither is a census of humanoid employees.
Their importance is less theatrical. The industrial base for deploying, maintaining and integrating robots is growing. China placed robotics at the center of its 2026–2030 industrial strategy; the IFR described that policy shift in May. In the latest global statistics, robot installations grew in China and the United States, while India reached roughly 10,500 installations, up 15 percent. Countries are not starting from identical levels, but the contest is wider than a simple China-versus-America race.
What matters most is not who produces the most impressive robot video. It is who can deploy equipment repeatedly, keep it working through normal production variation and teach the next task without starting the engineering process from zero.
A useful robot does not need to replace a worker
BMW’s account of its Figure 02 pilot in South Carolina is instructive. Over ten months, the company says the humanoid handled more than 90,000 sheet-metal components and accumulated approximately 1,250 operating hours while supporting production of over 30,000 BMW X3 vehicles. Its assignment was to position parts for welding in an already automated body shop. BMW also reports changes to safety arrangements and factory communications. This was a bounded deployment, not evidence that an entire assembly-line occupation had been replaced.
Warehouse automation offers a comparable example. In a commercial arrangement announced by GXO and Agility Robotics, Digit robots moved totes from existing mobile robots onto conveyors at a SPANX facility. The value proposition lay in a handoff between systems. The humanoid form was less important than whether the handoff could be performed reliably during the real shift.
Software progress may widen the range of viable tasks, but physical competence remains uneven. In its July announcement of Gemini Robotics 2, Google DeepMind reported 36 percent success at screwing in a light bulb and 92 percent at unscrewing one in a particular multi-finger dexterity evaluation. These are experimental results published by the developer, not deployment guarantees. Their contrast is a useful reminder that an apparently minor change in the operation can radically alter reliability.
The business calculation must therefore be the cost of accepted work over time, not the advertised robot price or the duration of a laboratory demonstration. Integration, maintenance, failed attempts, rescue by human employees, training, downtime and safety arrangements all belong in the denominator. A system that simply transfers labor from the production line to a group of remote troubleshooters has not removed the labor requirement; it has moved it.
The geography of production could shift before employment headlines do
These economics will not affect every country equally. Manufacturers with access to capital, engineering teams, spare parts, service providers and reliable infrastructure may find it easier to automate successive tasks. Others may remain competitive through skilled labor, proximity to customers and flexibility that machines cannot yet match. Lower labor costs also change the point at which automation becomes worthwhile.
The International Labour Organization’s work on robotics and reshoring warns against simple predictions of universal job loss or an immediate return of manufacturing to rich countries. Case studies found substantial technical bottlenecks and opportunities for human-machine complementarity. They also identified genuine exposure for economies whose advantage depends heavily on labor-intensive tasks that might be automated elsewhere. These results predate the latest physical-AI systems and are not a forecast of 2027. They show why the economic mechanism, rather than robot hype, deserves attention.
For emerging manufacturing countries, that creates a strategic question. They need not manufacture their own all-purpose humanoid to participate in the transition. They may need stronger integration skills, credible local testing, affordable financing, maintenance capacity and workers able to supervise, repair and improve automated processes. Investments that raise a supplier’s reliability can matter more than a nationally branded demonstration robot.
The labor market can change even if the headline number of employed workers initially remains stable. A factory might raise output while leaving vacancies unfilled, reorganize shifts or stop recruiting as many entry-level operators. Alternatively, higher output could sustain more jobs in complementary roles. Installation statistics alone cannot establish which outcome occurred. Governments that watch only layoffs may notice the change too late.
A better industrial policy measures the work
Public support for robotics should be tied to a practical evidence standard. A supported deployment should report, at an appropriate level of commercial confidentiality, the share of tasks completed without intervention, recovery time after failures, quality and scrap rates, safety performance and the full cost per accepted unit of work. The same test should run long enough to include ordinary breakdowns and production variation, not just a successful afternoon.
A second requirement concerns workers. Training subsidies should be linked to identifiable jobs, wage expectations and employers prepared to hire or promote people into them. Workers who already know a production process should help define the exceptions used to test automation. Their knowledge is a source of reliability data, not a courtesy to be collected after the technology is purchased.
Finally, industrial ministries should make testing and integration accessible beyond the largest manufacturers. Shared pilot facilities, interoperable interfaces and independent safety assessments could help smaller firms evaluate promising tools before signing expensive contracts. These programs should remain open to different robot designs. The aim is to develop an ecosystem capable of productive deployment, not subsidize whichever machine most resembles a person.
There is a legitimate counterargument: the cost of robotics may remain too high, and many jobs combine dexterity, judgment and unexpected situations that are difficult to automate. That caution is supported by present-day demonstrations. It does not eliminate the possibility of an important industrial change through much narrower uses. Several profitable task-level improvements, repeated across facilities, can matter without producing a single universally capable robot.
The world does not have to choose between announcing that humanoids will replace everybody and insisting that their limitations make the labor question premature. The decisive threshold is less dramatic: a task that was once too variable or expensive to automate becomes economical in a real workplace, under real constraints.
The countries best prepared for that moment will not necessarily be those with the most photogenic machines. They will be those that can measure usable output, adapt production and give workers a credible place in the transition. The robot workforce may arrive in the employment statistics long after it has begun to change the factory’s hiring plan.
