"We are seeing several signs that the 'ChatGPT-like moment' is nearing for physical artificial intelligence (AI) to enter our daily lives. The true 'ChatGPT moment' will come when robots can perform new tasks after seeing just a single example."
Abhinav Gupta, Skild AI co-founder and a professor in the computer science department at Carnegie Mellon University, said this about the 'inflection point for the spread of physical AI' in a written interview with ChosunBiz on the 14th. Skild AI is a leading unicorn in physical AI with a corporate value of about 20 trillion won, backed by investors including Jensen Huang, Nvidia chief executive officer (CEO), Masayoshi Son, SoftBank Group chairman, and Jeff Bezos, Amazon founder.
Gupta said robot AI currently needs tens of hours of data to perform new tasks, but the 'GPT moment' for physical AI will come when it can perform new tasks after seeing just one example, like humans.
Gupta, a world-renowned authority in Computer Vision and Robotics, said, "A new 'age of abundance' will open as AI-based robotic systems rapidly spread, centered on semi-structured spaces," naming factories, logistics centers, airports, and shopping malls as the 'semi-structured environments' where physical AI will first take hold. He said, "Semi-structured environments are more complex than spaces where traditional industrial robots operate, but the range of variables is more limited than in fully unstructured environments like homes, making them suitable for applying general-purpose robot AI in the real world," adding, "They will be the natural proving ground for physical AI."
Gupta majored in computer science and engineering at the Indian Institute of Technology Kanpur and earned a Ph.D. in computer science from the University of Maryland. As a professor at Carnegie Mellon, he has researched Computer Vision, self-supervised learning, and robot learning. He collaborated with Google on the scalability of Computer Vision and large-scale visual learning systems, and in 2018 joined Facebook AI Research (FAIR), expanding into large-scale AI research and Robotics.
Skild AI was founded in 2023, and its core technology is the Robot Foundation Model (RFM), which serves as the brain that determines a robot's actions. Trained on massive datasets, the model enables robots to autonomously explore, manipulate objects, and perform advanced interactions with their surroundings. Notably, it features generality that applies to all forms of robots, not only Humanoid Robots. Recognizing this, global corporations such as SoftBank and Nvidia moved quickly to invest. In April, LG Group Chairman Koo Kwang-mo visited Silicon Valley, and the two sides are working together. Gupta will deliver a keynote address at "SMARTCLOUD SHOW 2026," Korea's largest tech conference, at the Westin Josun Hotel in Sogong-dong, Seoul, on the 26th, speaking on the theme "Any Robot, Any Task, One Brain." The following is a Q&A.
◇ "No need to build separate AI for each robot"… Skild AI's 'general brain'
—What is the ultimate goal in developing general-purpose robot AI?
"Skild AI is building a foundation model for physical intelligence, that is, a general brain for robots. Our goal is to build a foundation model that enables robots to perceive and understand the physical world, reason about what they need to do, and predict what actions to take to achieve goals. The key is that a single AI must handle both a wide range of tasks and robot forms. The same intelligence should apply not only to humanoids but also to quadruped robots and robot arms. Put simply, our vision is 'Any Robot, Any Task, One Brain.'"
—You are pursuing a strategy of providing an AI platform rather than manufacturing robots yourself. What sets you apart?
"Skild AI's biggest differentiator is an 'omni-bodied' brain. Many robot AIs are closely tied to a specific robot form. Figure AI's Figure robot, Tesla's Optimus, and Physical Intelligence's mobile manipulator are the same. We believe this fundamentally limits robot performance. In physical AI, intelligence should not be bound to a particular robot body. Skild AI's goal is to build a brain that understands the fundamental principles of the physical world and interactions with it, regardless of specific forms such as humanlike robots, quadruped robots, or robot arms. AI should be able to control various robot forms—humanoids, quadrupeds, robot arms—and transfer intelligence across different tasks and environments."
—Many corporations are developing AI tailored to specific robots. Why is general-purpose robot AI more competitive?
"The biggest problem in robotics is data. When we talk about scaling robot data, it is not just about securing hundreds of millions or billions of data points; we need trillions. But collecting trillions of data points for just one robot form is practically impossible. Skild AI builds a foundation model for physical AI by consolidating data from all robot hardware, including humans, to build a general brain. Another important reason we need a general brain is safety."
—Does general-purpose robot AI also have advantages in terms of safety?
"Yes. In real-world settings, hardware issues can occur, such as motor failures or losing use of joints or limbs. AI optimized for a fixed, specific form can struggle in such situations, but an omni-bodied brain can assess remaining physical capabilities and change its behavior strategy. Even if a motor or limb fails, it can adjust its control strategy in milliseconds."
◇ Solving the data bottleneck with teleoperation, human videos, and simulation
—Data is the biggest bottleneck in physical AI. How do you address it?
"Unlike language, where large-scale internet data is relatively abundant and provides natural learning signals, robotics has a fundamentally different data problem. Robot data can be sourced from multiple streams with different strengths and weaknesses, such as teleoperation, human videos, and simulation. Teleoperation data is highly accurate and captures precise relationships among perception, actions, and physical outcomes, but it is expensive and limited in diversity, making it hard to scale.
Human videos excel in scale and diversity and are useful for learning general representations, but they lack information about actual robot motions. Simulation can generate large-scale data, but there is a sim-to-real gap. Therefore, Skild AI does not rely on a single data source and uses a staged data strategy. We pretrain models with video and simulation data, then train them on high-precision teleoperation data for real robot interactions and motions. This combines the scale and diversity of video and simulation with the precision of real teleoperation."
—You have investors such as SoftBank Group, Bezos Expeditions, Samsung, and LG. What drew their interest?
"Skild AI is not only building next-generation physical intelligence but also deploying it across a range of applications. Our robots are already being used in factories, data centers, warehouses, and even airports. Above all, Skild AI is made up of the strongest and smartest talent who have studied physical AI problems for more than a decade. My co-founder, Deepak Pathak, and I worked on these problems as robotics professors at Carnegie Mellon University (CMU). Our team has won multiple awards, including best paper awards at major robotics conferences."
◇ "Korea is a powerhouse in manufacturing and electronics… should partner with AI model companies"
—The Korean government aims to become a global leader in physical AI.
"Korea has a good opportunity to collaborate with companies in physical AI to develop its own robots and models. However, while it has traditionally led in manufacturing and electronics, it falls short in model building. By partnering with model development companies, Korean corporations can develop vertically integrated hardware and software products."
—Do you have plans to expand business in Korea?
"Korea is one of our target countries for expansion. We have already secured several customers, including major Korean corporations. We are currently driving automation of factories and logistics warehouses."