Headwaters Proposes "Physical AI Harness" Execution Platform and "SDI for AI Harness" Methodology to Safely Connect AI Decisions to the Physical World

Headwaters Inc. (Headquarters: Shinjuku-ku, Tokyo; Representative Director: Yosuke Shinoda; hereinafter "Headwaters"), a company engaged in the AI platform business, has announced its proprietary concept "Physical AI Harness" and its methodology "SDI for AI Harness" for the social implementation of Physical AI. This concept was publicly announced at "Physical AI Pitch Day," hosted by Microsoft Japan Co., Ltd. (Headquarters: Minato-ku, Tokyo) on June 18, 2026. Starting with the control of robot arms in manufacturing plant environments, the aim is to establish an execution platform that handles safety, explainability, and the structuring of on-site knowledge as an integrated whole.

Demo Videos:

- Robot Arm Picking Operation Demo Video 1

- Robot Arm Picking Operation Demo Video 2

Market Background: Physical AI is the "Next Battlefield," but there's a Wall Beyond POC Physical AI represents a turning point where AI moves from screens into the physical world, targeting the entire real economy, including manufacturing, logistics, and social infrastructure. The "AI Robotics Strategy" announced by the Ministry of Economy, Trade and Industry in March 2026 sets a goal of capturing a market scale of 20 trillion yen by 2040 (※1), and Fuji Keizai Co., Ltd. is conducting comprehensive market research on the global Physical AI-related market (※2). The Ministry of Internal Affairs and Communications' "Information and Communications White Paper 2025" also positions the development of the Physical Intelligence field as a factor supporting future AI market expansion (※3).

However, many cases in the field stop at the PoC (Proof of Concept) stage. Particularly in the BtoB manufacturing domain, implementation faces challenges such as: - Safety and explainability are prerequisites, as they can directly lead to process stoppages and accidents. - The acceptable level of defect rates is extremely strict, requiring quality at a "six sigma" level (99.99% is not enough). - The scope of existing automation is limited, and securing human resources, including at global sites, is a challenge.

These conditions mean that simply having a "smart model" is not enough to reach on-site implementation. The social implementation of Physical AI is shifting from a competition of model performance to a competition of implementation platforms.

Market Challenge: The Bottleneck is "Harness," Not "Model Intelligence" The primary bottleneck in practical application of Physical AI is the absence of peripheral infrastructure that "safely" and "explainably" connects AI decisions to physical world actions. In the field of Agentic AI, "Harness Engineering" is gaining attention as an environmental design philosophy for controlling AI that moves autonomously. This is positioned as an evolution of Context Engineering and refers to peripheral design that structurally suppresses runaway or erroneous operations while maximizing AI capabilities.

We are extending this concept to the physical world and view the "harness" layer in Physical AI as the core of social implementation.

Solution: Definition and Design Philosophy of Physical AI Harness We define "Physical AI Harness" as the entire peripheral system (sensors, actuators, low-level control, safety PLC, emergency stop mechanisms, simulation, etc.) for safely connecting AI to a physical execution system.

- Can be stopped before execution (the proposed action is approved/blocked/routed for human confirmation by the harness) - The reason for stopping can be explained (decision logs are stored in a structured knowledge graph) These are the design principles. While the implementation adopts a design philosophy combining neural (learning) and symbolic (rules), the true barrier to social implementation lies in "how to build the harness"—that is, how to translate on-site knowledge into a form that AI can handle.

True Differentiation: Proprietary Methodology "SDI for AI Harness" How to build the harness is a common issue for Physical AI entering the social implementation phase. In response to this question, we propose the methodology "SDI for AI Harness," starting with our proprietary engine "SyncLect Data Intelligence" (hereinafter "SDI") (※4).

SDI is a proprietary engine that converts tacit knowledge, such as on-site judgment criteria, exception handling, and preconditions, from business conversations, interviews, and documents into AI-ready structured data. We announced SDI in March 2026 and are providing vertical solutions for finance, manufacturing, and mobility.

SDI for AI Harness integrates the following sources from manufacturing sites and converts them into constraints that AI can handle (safety boundaries, termination conditions, reward design, etc.): - Extracting quantitative input conditions and normal completion conditions from work standards, QC process charts, and drawings.

- Extracting implicit prohibitions and recovery actions from expert interviews and on-site observations. - Quantifying thresholds and test requirements in a simulation environment.

Details: Synclect Data Intelligence Service Page

The business ontology and structured assets of tacit knowledge built by SDI form an imitation-resistant barrier and constitute the core differentiating element of Physical AI Harness.

Difference from Previous Approaches While AI has been applied to predictive maintenance and inspection before, the transition to the phase of autonomous judgment and action through "agenterization" has made an execution platform (harness) that ensures safety and explainability indispensable.

Demonstration: Robot Arm Control Assuming a Manufacturing Plant Environment At the presentation, a demo was conducted on the theme of robot arm picking operations in a simulated manufacturing plant environment. - AI proposes an action → Harness approves and executes it.

- Detects "dirt" on the object → Harness blocks and stops. - Requests human confirmation when necessary.

This realizes "moves intelligently, stops when dangerous, and reasons are visible" within a single loop. Demo videos can be viewed via the YouTube links above.

Future Outlook We will deploy this concept starting from manufacturing plant environments, our focus domain, and then expand it horizontally to logistics, social infrastructure, mobility, and other areas. - Design assuming process changes in manufacturing. - Standardization of safety and quality at global sites. - Collaboration with Microsoft and partnerships with other global partners. Based on these pillars, we will promote our business as a core technology company for social implementation.

Presentation Overview This presentation was given during the Partner Pitch slot (approx. 15 minutes) at "Physical AI Pitch Day" (approx. 90 offline attendees + online streaming), hosted by Microsoft Japan Co., Ltd. on June 18, 2026 (Thursday). Presenter: Ryuta Toshima, Connected Technology Department, IT Incubation Business Division.

※1: Ministry of Economy, Trade and Industry "AI Robotics Strategy" (announced March 2026). Sets a goal of capturing a 20 trillion yen market by 2040.

※2: Fuji Keizai Co., Ltd. "Physical AI Market Comprehensive Survey 2027" (Market Research Report).

※3: Ministry of Internal Affairs and Communications "Information and Communications White Paper 2025" Part II, Chapter 9, Trends in AI. ※4: Headwaters Announces "SyncLect Data Intelligence" to Convert Tacit Knowledge into Structured Data for AI Utilization.

End

Reference Information - Headwaters Fully Enters the Next-Generation Physical AI Market ~ Deploying "Autonomous Thinking Physical AI Technology" to the Enterprise Market with AI Agent Technology as the Core ~ https://www.headwaters.co.jp/news/headwaters_entering_physical_ai_market.html

- Headwaters Supports Physical AI Control of Autonomous Robots via Voice Commands Using Remote Robotics' "Remolink" ~ Realizing Collaborative Work Between Humans and Autonomous Robots On-Site Through a Design Where Humans Judge and AI Executes ~ https://www.headwaters.co.jp/news/Remolink_PhysicalAI.html

Trademarks Microsoft, Azure, and Microsoft OneLake are registered trademarks or trademarks of Microsoft Corporation in the United States and/or other countries. Other company names and product names mentioned are trademarks or registered trademarks of their respective companies.

Company Information Company Name: Headwaters Inc. Location: 4F, Shinjuku Island Tower, 6-5-1 Nishi-Shinjuku, Shinjuku-ku, Tokyo, 163-1304 Representative: Yosuke Shinoda, Representative Director Established: November 2005 URL: https://www.headwaters.co.jp/