Sam Altman, CEO of OpenAI, which developed ChatGPT, said this on the 1st (local time) as he moved to hire robot talent. He laid out a plan to expand AI, which had played multiple roles on screens such as answering questions and creating various text, images, and videos, into the brain of robots that operate in real life.
With Nvidia and other global corporations turning their attention to robots or physical AI one after another, the center of gravity in the AI market is shifting. Until now, AI-related investment and research have focused on infrastructure such as semiconductors and data centers, but recently the share of application areas such as Humanoid Robot is growing.
Investment bank (IB) Goldman Sachs said in a recently published report that "AI market investment has so far focused on securing graphics processing units (GPUs) and data centers," and projected that going forward, building the brains and bodies of robots that will move AI on industrial sites and elsewhere, and accumulating the data for that, will become more important.
AI models that serve as the brains of robots are already getting larger. Goldman Sachs assessed that the size of AI models for robots is expanding to the level of 40 billion to 80 billion parameters. Parameters are internal variables that AI adjusts during training. The larger the number, the more scenes, instructions, and motion data robots can learn, increasing their ability to respond to complex situations. Robot AI, too, is growing from small lab-stage models to sizes closer to large language models (LLMs).
Accordingly, robot technology is moving from a stage of simply following commands to a stage of predicting the results of actions in advance and moving. This combines a world model, which calculates the outcome of movements in advance, with a vision-language-action (VLA) model that has robots view their surroundings with cameras, understand human instructions, and convert them into actions.
In areas where contact is crucial, such as tasks that grasp or insert objects, there are also efforts to boost force control capabilities by using a vision-touch-language-action (VTLA) model that adds tactile information. To acquire these capabilities, sufficient real-world behavioral data—such as recognizing, grasping, and moving objects—must be accumulated.
Because it takes time to accumulate data in real situations, the robotics industry expects large-scale commercialization to begin in earnest between next year and 2029. Current robot applications are centered on factories and logistics transport work.
An industry official said, "Right now, we are at the stage of deploying Humanoid Robot by the dozens to logistics warehouses and manufacturing sites to check operational stability and data quality." The official added, "With only simple and repetitive field tasks, it is limited to accumulate data to advance robots, and moves to produce and process diverse data in separate data factories are accelerating."
Investment is also concentrating on key supply chain businesses needed for robot commercialization. Goldman Sachs cited harmonic reducers and actuator modules as representative key components. Harmonic reducers determine the precision and torque of robot joints. Actuator modules bundle motors, reducers, sensors, and controllers to power the arms and legs of robots.
Goldman Sachs said in the report, "Labor shortages and demand for automation are structurally accelerating robot adoption," and assessed that "the robotics industry has entered the early cycle of a long-term capital rotation."