Karakuri, JTEKT, and Upstage Jointly Selected for "Physical AI Development Support Program by AWS Japan"

Karakuri Inc. (Headquarters: Chuo-ku, Tokyo; CEO: Shimon Oda; hereinafter "Karakuri"), a company developing domestic multimodal AI, in collaboration with JTEKT Corporation (Headquarters: Kariya City, Aichi Prefecture; President: Yoshihito Kondo; hereinafter "JTEKT"), a key company of the Toyota Group, and Upstage AI Inc. (Headquarters: Minato-ku, Tokyo; Representative Director: Hiroyuki Matsushita; hereinafter "Upstage"), a global AI company, announce their selection for the "Physical AI Development Support Program by AWS Japan" provided by Amazon Web Services Japan G.K. (hereinafter "AWS Japan").

Through the adoption of this program, the three companies will verify the potential of physical AI in the manufacturing industry and aim for the social implementation of autonomous and adaptive production systems in JTEKT's factories.

Background of Selection and Significance of This Project

While Japan's manufacturing industry boasts world-class technological capabilities, it faces structural challenges such as a severe labor shortage and the transfer of skills from experienced workers. In particular, the burden of "teaching" (advanced programming) in robot control is significant, hindering flexible operations on the factory floor. In this project, by combining Karakuri's "LLM/Advanced Intent Understanding," Upstage's "Data Structuring Technology to Enable AI Utilization of Unstructured Data," and JTEKT's "Extensive Knowledge Cultivated in Manufacturing Sites," we aim to implement Japan-originated physical AI that allows intuitive robot operation through "words" and "visual information" even without specialized knowledge.

Why Karakuri Enters Physical AI – Technology Transfer from "Customer Support" to "Physical AI"

The background to Karakuri's entry into the physical AI domain is the high technical affinity between the Computer Operation Agent (CUA) model it developed for customer support and the Vision-Language-Action (VLA) model used for robot control.

CUA Model Mechanism: It recognizes images on the screen and autonomously generates actions such as "where to move the mouse and what to click" from natural language instructions from the user.

Commonality with VLA Model: The process of "seeing an image (Vision), understanding it with language (Language), and generating an action (Action)" is essentially the same structure as the CUA model.

This technology transfer accelerates the development of physical AI in the following three key areas:

Three-Company Collaboration System and Each Company's Role

In this project, the three companies, each with their strengths, will collaborate to promote the social implementation of physical AI in manufacturing sites.

Joint Development by the Three Companies

Utilization of Vision Language Action (VLA)
Multimodal generative AI integrates learning and inference of images, language, and actions to achieve automation and optimization of complex manufacturing processes.

Reduction of Prototyping Costs through High-Speed Learning in Virtual Space (Sim2Real)
By transferring learning outcomes from simulation environments to real environments, AI model development costs and prototyping periods are significantly reduced.

Assetization and AI-Readiness of On-Site Data
Diverse and highly accurate data accumulated in JTEKT's factories will be converted into a format that AI can maximize, and dedicated models based on on-site data will be built.

About the "Physical AI Development Support Program by AWS Japan"

The "Physical AI Development Support Program by AWS Japan" is a program implemented by AWS Japan for companies and organizations with legal entities or bases in Japan. It provides technical support, AWS credits (up to 6 million USD for the entire program), formation of a physical AI community, and Go-to-Market support for the development of robot foundation models including VLA. The support period is approximately six months from March 2026, with a results presentation scheduled for July.

URL: https://aws.amazon.com/jp/blogs/news/aws-japan-physical-ai-development-support-program/

Comment from Toru Onozaki, Head of Research and Development Center, Innovation Division, JTEKT Corporation

To promptly respond to the decline in the working-age population due to Japan's declining birthrate and aging society, I feel that it is now essential to elevate the "master craftsmanship" and "on-site knowledge" cultivated by JTEKT into physical AI.

In the manufacturing sites of automotive parts, bearings, and machine tools, a vast amount of tacit knowledge, backed by the experience, intuition, and tips of skilled technicians, has been accumulated. However, much of it has not yet been sufficiently formalized even with the power of digital technology.

By combining Upstage's Document AI technology and Karakuri's generative AI technology, we aim to convert on-site knowledge into explicit knowledge that AI can handle, thereby accelerating the implementation of physical AI in manufacturing sites.

We position this initiative as the first step in redefining Japan's world-class manufacturing competitiveness and will co-create a model for physical AI implementation in the manufacturing industry.

Comment from Hiroyuki Matsushita, Representative Director, Upstage AI Inc.

The manufacturing industry operates on data, but much of it is still trapped in documents, logs, and fragmented systems. Upstage's strength lies in converting such unstructured information and tacit knowledge dormant in the field into explicit knowledge and intelligence that AI can utilize. By combining JTEKT's manufacturing expertise, Karakuri's generative AI development capabilities, and Upstage's Document AI technology, we believe this is an opportunity to implement physical AI from concept to the manufacturing floor.

Comment from Tomofumi Nakayama, CPO, Karakuri Inc.

The strength of Japanese manufacturing lies in its "on-site capabilities." It is the ability to solve complex problems that cannot be conceived from the top down, by iterating improvement cycles on the factory floor. To leverage this strength in the age of AI, a mechanism that allows on-site personnel to intuitively use AI is indispensable. The technology of "understanding the intent of human language and responding appropriately," refined in the customer support domain, can be directly applied to controlling robots with human language. By partnering with Upstage's data structuring technology and JTEKT's world-class manufacturing sites, we aim to demonstrate a new form of Japan-originated physical AI.

Company Overviews

JTEKT Corporation

Location: 1-1 Asahi-cho, Kariya City, Aichi Prefecture

Representative: Yoshihito Kondo, President and Representative Director

Business Activities: Development, manufacturing, and sales of steering systems, driveline components, bearings, machine tools, electronic control devices, etc. Holds the world's No. 1 market share in automotive steering systems.

URL: https://www.jtekt.co.jp

Upstage AI Inc.

Location: Minato-ku, Tokyo

Representative: Hiroyuki Matsushita, Representative Director

Business Activities: Provision of safe and high-performance LLM and Document AI platforms. Develops enterprise AI platforms such as the domestically developed LLM "Syn Pro" and Document Parse, with investments from Amazon and AMD.

URL: https://upstage.ai

Karakuri Inc.

Location: Camel Tsukiji II 5F, 2-7-3 Tsukiji, Chuo-ku, Tokyo

Representative: Shimon Oda, CEO and Representative Director

Established: October 2016

Business Activities: Development, provision, and operation of the "KARAKURI" series of AI specialized for customer support, and research and development of large language models (LLMs).

URL: https://karakuri.ai