Exclusive | Taiwan Built the Chip World. Robots Could Be Next.

As humanoid robots move from research labs to factory floors, a leading MIT artificial intelligence researcher says Taiwan's manufacturing depth and semiconductor supply chain give it a rare opening to anchor the global robotics industry — provided the island is willing to bet on software as well as hardware.

​Pulkit Agrawal, a professor of electrical engineering and computer science at the Massachusetts Institute of Technology and a specialist in robotics, deep learning, and computer vision, shared that assessment during an exclusive interview with Storm Media in Taipei in late June. Agrawal had traveled to Taiwan to participate in the Epoch Foundation and MIT Global Industry Research Program Taiwan Annual Forum and Technology Forum. A doctoral graduate of the University of California, Berkeley, he has also co-founded two robotics startups — Safely You and Eka Robotics — that translate research into commercial applications.

His visit coincided with growing international attention on physical robotics. NVIDIA CEO Jensen Huang has said publicly that embodied AI — robots that perceive and act in the physical world — represents the next major wave of artificial intelligence, and that Taiwan's supply chain is at the heart of the infrastructure required to make it real.

Agrawal endorsed that view without reservation. "Taiwan has always been a leader in manufacturing," he told Storm Media. "The robotics industry requires an enormous amount of hardware manufacturing — and this is a tremendous opportunity for Taiwan. These robots need massive computing power, and Taiwan has a complete supply chain in semiconductors, computers, and data centers. Taiwan will play a very important role in building the robotics ecosystem."

In his analysis, two areas give Taiwanese companies a structural edge: manufacturing the physical components of robotic systems, and supplying the computing infrastructure — chips, servers, and data centers — that runs the AI models controlling them.

But Agrawal was careful to define where that advantage hits its ceiling. "Taiwan has traditionally been manufacturing-oriented," he said. "If it can move toward software and invest heavily in the software industry, it will have even greater future potential." The observation carries weight at a moment when robotics competition is increasingly decided by algorithms and data pipelines, not just precision components.

The commercial imperative for advanced robotics is, in part, demographic. Across much of Asia — including Taiwan, Japan, and South Korea — falling birth rates and aging populations are shrinking the labor pool faster than immigration policy can compensate. Agrawal sees robots filling specific gaps rather than replacing the workforce wholesale.

"Modern society faces many challenges," he said. "An aging population, many dangerous jobs — cleaning sewage systems, mining — and many highly repetitive, tedious jobs like washing dishes and cleaning floors. Robots can take on these roles and become part of the solution to many problems."

For elderly people who wish to live independently, the use case becomes both economic and humanitarian. Agrawal described a near-future scenario in which household robots handle cleaning, cooking, and basic upkeep, and where robot companions — robotic dogs and cats, for instance — provide forms of social and cognitive stimulation that help older adults stay mentally active and reduce the risk of dementia. "In the next 10 to 20 years," he said, "robots could take over many human jobs or even surpass human capabilities in certain areas."

Despite that long-run confidence, Agrawal was precise about where the technology falls short today. "Robot arms are still very far from achieving everything a human arm can do," he told Storm Media. "A robot arm can perform delicate tasks — tightening a screw, for instance — but the motion is still unstable. It might succeed 10 out of 20 attempts, but not every time."

Multi-task generalization — the ability to pick up a watch, repair it, brew coffee, and clean a room in sequence — remains out of reach. "We are still very far from that ideal," he acknowledged. "It requires upgrades to both hardware and software."

A less obvious but equally fundamental challenge is what Agrawal calls "force intelligence": the calibrated physical feedback that tells a robot arm how tightly to grip a coffee mug — enough to hold it without dropping it, not so much as to shatter it. Achieving that kind of tactile judgment at scale, he argued, is one of the most underappreciated problems in the field.

Data scarcity compounds the difficulty. Unlike large language models, which could train on the vast text already available on the internet, robotics AI requires physical motion data that does not exist online at scale. Researchers are addressing this two ways: through teleoperation — humans wearing sensor suits perform tasks while robots record and replicate the movements — and through digital simulation environments such as NVIDIA's Omniverse, where robots practice in near-realistic virtual worlds before encountering the real thing. Whether performance in simulation reliably transfers to real environments, Agrawal noted, remains an open research question.

His own lab is working on a different front: core algorithms that let robots switch between tasks rapidly, without months of retraining. "Currently, robots need to be pre-programmed by human engineers — a process that can take three months, six months, even a year to train a robot to complete one task," he explained. "We are working hard to develop core algorithms that help robots switch task modes quickly. Think of it like an iPhone receiving a software update — new functions are immediately available."

The longer-term vision is more ambitious still: giving robots the capacity to learn from video demonstrations and accept instructions in plain language. "Our goal is to command a robot with natural language to clean the floor or do household chores," Agrawal said. "You could play a video and ask a robot to watch it and learn a new task on its own — just as you can ask ChatGPT anything, one day you will be able to direct a robot to do a wide range of things."

The economic upside of robotics is inseparable, in Agrawal's view, from its social risks. In manufacturing, logistics, food packaging, hospitals, hotels, and restaurants, he sees major commercial opportunities: robot-operated factories can run 24 hours a day, 365 days a year, in the dark, with significant energy savings — and can scale production rapidly when demand surges, as the semiconductor industry knows well.

But large-scale automation also means large-scale displacement. "Many dangerous, repetitive, and tedious jobs can be handled by robots," he said. "And for those who become unemployed as a result, government, industry, and academia must work together on solutions — deciding which environments robots should be placed in, providing retraining and education for displaced workers so they can find new jobs and sustain their livelihoods. This point is extremely important."

He framed the challenge in systemic rather than technological terms. "Robots are not competing with humans," Agrawal said. "Automation is there to help humans accomplish tasks. Once you place a robot in an environment, you have to think about the entire ecosystem — the impact on people, and how to help displaced workers maintain their livelihoods."

On governance, Agrawal called explicitly for cross-sector collaboration rather than leaving regulation to any single stakeholder. "Technology cannot develop in isolation from society," he said. "It must be shaped by conversations among academics, AI developers, and government decision-makers. Scientists can help business leaders and policymakers make the best decisions — to advance society, rather than letting technology strip away the things we value." Some decisions, he added, must always remain with humans: wherever ethics, physical safety, or life-and-death consequences are involved, the final control must stay in human hands. （Related： CarbonSix Secures $40M Series A to Deploy Physical AI Across Global Manufacturing ｜ Latest ）

​Original Article in Chinese​