On July 2, 2026, I visited the "Monozukuri World Physical AI Exhibition" being held at Tokyo Big Sight. The evolution of generative AI and reinforcement learning is breaking out of the digital world and exploring how to intelligently operate robots in the real world (physical). This exhibition was a very vibrant space that truly hinted at the next generation of manufacturing and the automation of R&D (research and development). I will report on the latest trends that particularly caught my eye on-site!

Now, AI is trying to leap from "the screen" into "the real world." Local LLMs, which excel in security and high-speed processing, have now become a powerful practical engine that goes beyond mere dialogue tools, playing a central role in corporate DX and products. Many of you may have seen examples like "life-sized AI avatars for customer service" being introduced in convenience stores. What happens when local LLMs are further integrated into hardware such as communication robots? AI will evolve into "Physical AI," capable of recognizing the real world and physically interacting with it. The key to this paradigm shift lies in cutting-edge technology known as the "VLA (Vision-Language-Action) model." In this context, we have created an experimental prototype to actually experience the "sense of AI connecting with the real world." We would like to introduce you to its operation and potential!

What is the key to dramatically improving development speed in R&D? The R&D (Research and Development) in manufacturing and mobility development is currently undergoing a significant transformation. With the increasing performance and complexity of products, the amount of data and the computation time required for simulations have exploded. In traditional CAE, even with the use of high-performance HPC, it was necessary to spend enormous computational resources and lead time to obtain high-precision analysis results, taking "days to weeks for a single computation," which has become a major bottleneck in the design cycle. In this "AX (AI Transformation) era," the key to dramatically accelerating the speed of research and development is the "Surrogate Model AI." This article will explain the technological breakthroughs that AI-driven simulations bring to real design and the specific approaches for their implementation.

In the manufacturing industry and advanced research fields, the use of real machines for trial and error is already reaching its limits. NVIDIA's "Physical AI Platform" aims to completely eliminate the heavy constraints of "time, cost, and risk," elevating development speed to an entirely new dimension. This approach, known as "Sim2Real," allows for safe and instantaneous learning in virtual spaces, conducting tens of thousands to hundreds of millions of iterations, and seamlessly implementing the results into the real world. This is not just a concept; it is a powerful ecosystem that has already been put into practical use. In this article, we will elucidate the three phases that make up this platform and the roles of specific libraries that hold the key to its implementation.

The development of autonomous driving systems is transitioning from traditional rule-based and modular approaches to an end-to-end "AX (AI Transformation)" environment where AI comprehensively understands situations and makes decisions. NVIDIA's offerings, "Alpamayo," "AlpaSim," and "NuRec," represent the latest ecosystem aimed at the societal implementation of level 4 autonomous driving, balancing safety and explainability. NVIDIA's autonomous driving AI platform "Alpamayo" moves away from reliance on traditional pattern recognition in autonomous driving by incorporating a mechanism where AI "infers (thinks and judges)" like a human, dramatically enhancing safety and reliability. The R&D departments of enterprises will report on the HPC requirements necessary to maximize the use of these technologies, integrating physics-based simulations (CAE) with AI models.