"Just as ChatGPT gives an immediate answer when you ask a question, now we have entered an era where robots can cook right away without pretraining if you ask, 'Make me pancakes,' and show a related video. The 'ChatGPT moment' has come to the field of physical AI as well."
Abhinav Gupta, co-founder of Skild AI and a professor in the Department of Computer Science at Carnegie Mellon University, said on the 26th at the Westin Josun Hotel in Jung District, Seoul, at SMARTCLOUD SHOW 2026, the country's largest tech conference, "Until now, to make a robot perform a new task, we had to collect a large amount of related data and then do pretraining, but now, by showing a single video, it can carry out tasks it has not done before."
Skild AI is a global physical AI company that develops a Robot Foundation Model (RFM), the 'brain' of robots. It has received investment from Nvidia, SoftBank, Amazon, LG, and Salesforce Ventures, and its corporate value has been estimated at about $14 billion (about 20 trillion won).
In a keynote titled "Any Robot, Any Task: One Brain," Gupta said, "The biggest bottleneck in Robotics is not robot hardware but data," adding, "In the case of Generative AI, it has rapidly improved performance and scaled by training on vast text and image materials available online, but in physical AI there has been a lack of high-quality behavioral data for robots to learn from, so progress has not been as fast."
He emphasized that to build a general-purpose robot brain not limited to specific robot models or task types, Skild AI has used every kind of robot data collection method and, as a result, has secured 100 to 1,000 times more data than competitors.
Gupta said, "Skild AI's RFM is pretrained with video data and synthetic data from virtual simulations, and it undergoes subsequent training and fine-tuning through teleoperation, in which a person remotely adjusts a robot's actions, and on-robot field learning," adding, "Through this, we have implemented a robot brain that can autonomously explore its surroundings, manipulate objects, and perform precise tasks."
Robots equipped with Skild AI's RFM have been deployed to real industrial sites and are performing difficult tasks previously handled by people. Gupta said, "At Nvidia's Houston plant, the deployed robots can perform precise work such as tightening graphics processing unit (GPU) screws," adding, "At a Foxconn plant, robots assembled more than 16,000 GPUs in a month, raising the factory's level of automation."
He cited as a differentiator that Skild AI's RFM has generality that applies not only to humanoid robots but to all forms, including bipedal, quadrupedal, and wheeled types. Gupta said, "The key is building an omni-bodied brain that controls a variety of tasks performed by all forms of robots in multiple locations, including indoors, outdoors, factories, and data centers," adding, "Because other companies' robot brains are tailored to specific robot hardware, they stop operating when a failure occurs or an unexpected situation arises, but Skild AI's robots detect changes when the hardware is switched and continue the work."
Skild AI on the day unveiled S1, a general-purpose model in which a robot immediately learns a new task and executes it after watching just one video. Gupta said, "It is the first model designed so that, without training on the new task, the robot learns previously unseen tasks using only video prompts," adding, "This is a significant advance in that it resembles how a person observes another person's actions to learn something new."
He added, "As robots' generalization capabilities strengthen in this way, we expect the 'ChatGPT moment' to arrive soon in the field of physical AI."