The global robot large language model (robot LLM) market, including AI training and AIGC solutions, is expected to exceed US$100 billion in 2028.
According to the latest research, as humanoid robots move towards a high degree of system integration and are expected to move from industrial scenarios to family life, front-end AI model training will become more critical to meet more back-end understanding and interaction needs. It is estimated that the global robot large language model (robot LLM) market, including AI training and AIGC solutions, is expected to exceed US$100 billion in 2028, and the annual compound growth rate from 2025 to 2028 will reach 48.2%. NVIDIA launched the Cosmos platform with a built-in World Foundation Model (WFM) at CES 2025, which is expected to become an important market driver.
The Sim-to-Real Gap is a major challenge in robot training. Since deep learning and neural networks are simulated and learned in a virtual environment, robots may ignore real-life physical properties, sensor information, or actuator power, thereby reducing their actual operating efficiency. Therefore, it is necessary to strengthen technical tools that simulate the real world. The NaVILA model released by NVIDIA in December 2024 allows robots to adapt to complex terrain and plan paths using visual systems and LiDAR without the need for training and maps, as long as they understand natural language commands.
Analyzing the current cost of humanoid robots, AI training and AIGC solutions account for the highest 10% of the software side, while the rest of the operating system accounts for about 4%, and automatic navigation accounts for 3%. Taking the NVIDIA ecosystem as an example, Siemens, Vention, Hexagon and other manufacturers have invested in the field of robot simulation, while Rockwell Automation, PTC and Cadence have invested in computer-aided engineering (CAE) and reality capture, and Taiwan's new startup MetAI is also one of them.