Free Lecture》Humanoid robots can work, which is the industrial revolution

South Korea recently approved twenty bipedal humanoid robots for field testing in semiconductors, automobiles, shipbuilding and logistics. Twenty units is not too many, but it is worth noting that the South Korean government looks at humanoid robots from a different perspective, making factories not only a place for using robots, but also a place for training robots.

This batch of robots will clean semiconductor parts, carry automobile body components, stack shipbuilding scaffolding, and carry cold chain packaging materials in real production lines. Some tasks will even share or be adjacent to the same working space as humans. The South Korean government plans to use these tests to accumulate safety and operational data, and then go back to revise the current safety regulations for industrial robots. This is no longer simply purchasing equipment, but using factories to cultivate technology and using on-site systems to improve systems.

With the development of humanoid robots today, the most difficult thing is no longer to make it stand up, take a few steps, or even let its arms grab things. The real threshold is how to make robots continue to work in uncertain environments. What should I do if the parts are out of position? What should you do if someone suddenly passes by? How to correct the wrong grabbing power? Can the task still be completed after the workpiece specifications are changed? These problems can no longer be solved solely through laboratory simulations.

And this is where behavioral data starts to get expensive. The angle at which skilled workers hold parts, how to turn their wrists, how to stop when they see an abnormality, and how to adjust the same work under different circumstances were all regarded as the master's experience in the past, but now they may become training materials for physical AI.

The scale of South Korea's humanoid robot industry currently lags behind that of the United States and China, but South Korea has large manufacturing sites for semiconductors, automobiles, shipbuilding, and batteries, as well as a large number of skilled technicians. What it wants to leverage is converting existing manufacturing capabilities into learning materials for humanoid robots. If we look at it from this perspective, there are actually many places that Taiwan can learn from.

Taiwan has the highest density of advanced manufacturing sites in the world. From wafer manufacturing, packaging and testing, PCB, servers, electronic assembly, to machine tools and precision machinery, we have no shortage of factories, engineers and technicians. But what is really lacking are institutional arrangements for how to turn these advantages into next-generation robot capabilities. When talking about humanoid robots in Taiwan, attention often focuses on reducers, motors, ball screws, joint modules and controllers. Of course these are important, but if we only supply components, we will eventually be able to return to Taiwan's familiar foundry position; then we will make the hardware, others will get the data, others will train the models, and the platform specifications will be determined by others.

But in the AI ​​era, what we should really grab is industrial behavior data. Taiwan should start from semiconductor back-end, electronic assembly, warehousing logistics, equipment inspection and high-risk operations, select tasks with clear work boundaries and controllable risks, and let robots accumulate hundreds or thousands of hours of data in real sites. Successful actions should be retained, and failed actions should be retained; how many times manual intervention is required, what situations are likely to fail, and which actions require correction by skilled workers, these should all become training materials for the model.

This also means that the government cannot just subsidize companies to buy robots. What is important is to establish a regulatory sandbox where humans and machines co-exist, so that factories dare to test, manufacturers can make changes, and data can be accumulated. Subsidy indicators should not only look at the number of robots purchased, but should also require autonomous operation rate, number of manual interventions, task success rate, mean time between failures and the cost of each successful task. If the policy still uses equipment procurement thinking to promote physical AI, it may only increase the number of robot demonstration cases in the end, but will not increase the industrial capabilities of humanoid robots.

The competition among humanoid robots in the future will not only be about whose robot moves the most like a human, but who can make the robot learn to work the fastest. China has the scale, the United States has the models and capital, and South Korea is turning its factories into data training grounds. Taiwan's advantage is not just components, but one of the few high-density advanced manufacturing sites in the world. If we only think about what robot parts Taiwan can supply, we will underestimate the problem. The more important question should be whether Taiwan can transform decades of accumulated manufacturing experience in factories into data assets that are the most difficult to replicate for next-generation robots?