Humanoid Robot is a moving, giant semiconductor system. Physical artificial intelligence (AI) cannot secure competitiveness with only a single high-performance chip.
Park Jun-sik, head of STMicroelectronics Korea, said this in a presentation titled "Physical AI: Why robots must sense, judge, and move like humans" at SMARTCLOUD SHOW 2026, the country's largest tech conference, held at the Westin Josun Hotel in Sogong-dong, Seoul, on the 26th.
Physical AI refers to technology that connects AI's recognition and judgment of the real world to the actual movement of machines. Park said, "In the era of physical AI, the ability to take the entire system—from sensors, microcontrollers (MCU), analog semiconductors (semiconductors that process sensor and power signals), motor control, to Power Semiconductor—through to mass production is important," adding, "ST supports almost every area in semiconductors except the brain."
Park cited sensing as the first stage of physical AI. He said, "Just as people do not judge everything with a single eye, robots are operated through sensor fusion (technology that combines data from multiple sensors), including eyes, ears, and touch, as well as the sense of balance," and added, "Not only AI processors but a wide range of sensor technologies are coming to the fore."
He presented the humanoid as the representative platform for physical AI. Park said, "Human environments are built around humans," adding, "Robots with a humanlike form can perform a variety of tasks without changing existing environments." As AI's operating range expands from the cloud to the real world, humanoids are drawing attention as a representative implementation of physical AI.
Park said, "In physical AI, AI goes beyond simply analyzing information and performs actual physical actions," adding, "The process of sensing, perceiving and understanding, judging, and acting is repeated in real time within a single system." He said physical AI is evolving from cloud AI that runs models in large data centers, through edge and embedded AI (AI that runs near or inside the device) in smartphones, cars, and industrial equipment, to AI that moves directly in the real world.
Park also emphasized that for a humanoid to move naturally, not all decisions should be left to a central AI processor. Sending sensor data to the center increases traffic and latency, so decisions must be made immediately on site, he said. "If you give the command 'Pick up that cup,' how much to move the fingers and how much current to supply to the motor is handled by the MCUs around each joint," he said. "The AI processor thinks, and the MCU immediately adjusts the movement." He added, "It's similar to how the brain does not calculate every minute movement of every muscle, but the nervous system and reflexes handle fast actions," and said, "As humanoids advance, distributed intelligence architectures will become even more important."
He also cited "closed-loop control" as a core technology, in which the robot checks the result with sensors after moving and corrects its motion. With closed-loop control applied, a robot does not stop at executing a command once; it measures the current state with sensors, compares it with the target value, and adjusts the motor's speed and torque in real time.
Citing analyses by global investment banks (IB) such as UBS, Goldman Sachs, and Morgan Stanley, Park projected that humanoid production will increase from about 18,000 units last year to about 256,000 units in 2030. ST estimates the bill of materials (BOM) value for semiconductor components it can address in a single humanoid at $600 (about 830,000 won).
Park said, "As the humanoid market grows, demand will inevitably rise simultaneously not only for AI processors but also for embedded AI (AI that runs inside devices), analog semiconductors, sensor ICs, and power semiconductors."
Park also presented cost-performance, regulation and social acceptance, standardization, return on investment (ROI), and safety and cybersecurity as conditions for mainstream adoption of humanoids. He said, "In physical AI, cybersecurity issues can expand from an information issue to a physical safety issue," adding, "Mass adoption will be possible not when a single technology is perfected, but when technology, business, and the overall ecosystem are ready at the same time."