By VADM David Lewis (USN, Ret.), President of the American Society of Naval Engineers

AI is moving deeper into authoritative engineering data rather than remaining a conversational add-on, and maritime AI is beginning to acquire the systems-engineering and assurance structures needed for real deployment. The most important development, therefore, is not “AI is getting better at CAD.” It is that the surrounding engineering system is becoming ready for AI.

Structured BOMs are becoming conversationally addressable. PMI can travel directly into machine inspection. APIs expose CAD, drawings, and manufacturing simulation to agents. Ship designers are beginning to demonstrate AI inside their domain tools. Classification societies are engineering approval pathways for AI-assisted ships. Reality capture can reconcile twins against the physical asset. Physical AI is moving into variable industrial environments. Metrology is becoming AI-assisted.

That is the transition from a copilot sitting beside engineering to an agent operating inside a governed engineering-production system.

For naval shipbuilding, that distinction should guide investment. The highest-value program is unlikely to be an isolated “Navy generative-AI CAD tool.” It is the creation of a machine-addressable, model-based, configuration-controlled engineering and manufacturing enterprise in which increasingly capable agents can safely operate from requirements through design, production, inspection, sustainment, and eventually fleet feedback.

The next major checkpoints come at IMTS in Chicago, 14–19 September, where Siemens and FANUC are preparing significant Industrial AI, CNC, digital-thread, physical-AI, and virtual-commissioning demonstrations, and at ICCAS in Singapore, 14–16 September, where NAPA and its partners will address design-to-operations digital twins and 3D model-based ship classification. 

The direction is increasingly:

AI agent → authoritative product data → model-based definition → deterministic engineering tools → production/inspection → verified physical state.

That is a considerably more mature proposition than simply asking an LLM to generate a 3D model.

The developments below fall into three connected layers: authoritative digital engineering, governed maritime implementation, and closed-loop production and sustainment. The first layer covers product data, naval-architecture models, classification, and edge deployment; the second addresses systems engineering, human oversight, and industrial adoption; the third connects robotics, metrology, reality capture, and physics-grounded optimization. Reading the briefing in that sequence clarifies how individual announcements contribute to one engineering-production architecture.

1. Autodesk Fusion gives AI authority over product data—not just CAD commands

Autodesk’s 3 September Fusion update is the most consequential mainstream CAD/CAM release of the week. Autodesk Assistant is leaving Tech Preview and becoming a generally available Fusion capability. More importantly, it can now find components and perform bulk property changes through natural language, while Fusion Manage users can create, modify, delete, and populate custom property definitions conversationally. 

At the same time, Fusion added direct authoring of Product Manufacturing Information. Hole/thread notes can carry diameter, depth, tolerance, thread, counterbore, and countersink information under ASME or ISO conventions. Downstream probing operations can inherit asymmetric tolerances from that PMI rather than requiring manual re-entry. 

Two quieter changes matter just as much for agentic engineering. NETROUTE is explicitly designed for scripting, automation, and AI-driven schematic generation, including an unattended mode; and Fusion’s API now expands automation across modeling, drawing creation, and process simulation. Autodesk also moved its Metal Powder Bed Fusion Process Simulation API out of preview, enabling programmatic execution and result retrieval. 

Why it matters: the Assistant is moving from helping an engineer use Fusion toward acting upon the structured engineering definition itself. That is a critical distinction. Consider the potential chain:

natural-language engineering instruction → BOM/property change → PMI → CAM/probing → simulation → inspection evidence.

For defense manufacturers, this is the architecture needed for AI-assisted model-based engineering. The agent can manipulate controlled engineering objects while the CAD kernel, standards, simulation engines, CNC controllers, and inspection systems continue to provide deterministic constraints.

One particularly interesting naval example is Fusion’s new two-axis probing alignment for rotational components such as propellers and impellers—exactly the sort of high-value, low-volume machinery where connecting model definition, machining, probing, and inspection can yield substantial benefit. 

2. NAPA demonstrates AI inside actual naval-architecture software

At SMM Hamburg on 1 September, NAPA demonstrated NAPA Navigator, describing the presentation as “AI-Powered Productivity in Ship Design.” NAPA later said its engineers showed attendees how Navigator uses AI to accelerate ship-design work. This is still being presented as an emerging capability rather than a fully documented production release, so treat performance claims accordingly. 

The context around Navigator may be even more significant. NAPA announced work it will present at ICCAS later this month on connecting design-stage and operational data into multidimensional vessel digital twins, and on the OCX standard’s role in moving classification from conventional drawings toward 3D model-based review. The latter effort includes DNV, Bureau Veritas, Lloyd’s Register, PROSTEP, and Cadmatic. 

Why it matters: this is one of the clearer signs that AI is penetrating domain-native naval architecture, rather than reaching ship designers only through generic engineering copilots.

The larger opportunity is not an AI that draws an entire ship. It is an agent able to operate within a naval-architecture environment containing hull form, hydrostatics, stability, structure, arrangements, rules, calculations, and established design procedures.

Combine that with a model-based classification standard and a design-to-operations digital twin, and the eventual workflow becomes much more powerful:

design intent → naval architecture model → class rules → analysis → production information → as-built ship → operating data → next design iteration.

That would make the AI part of the ship’s lifecycle engineering system, not merely an aid to the designer.

3. CMA CGM, SDARI, and Bureau Veritas put AI-assisted ships through formal systems engineering

A particularly important SMM announcement came on 3 September. Bureau Veritas, CMA CGM, and Shanghai Merchant Ship Design and Research Institute launched a Joint Development Project to develop an assisted container-vessel concept using AI, digitalization, and enhanced decision-support functions. 

The project's methodology is noteworthy. It will identify onboard functions appropriate for different levels of assistance, develop a Concept of Operations and Basis of Design, define crew, automated-function, and shore-support roles, conduct safety and human-factors assessments, and ultimately develop technical specifications and seek Approval in Principle. The team will also assess CAPEX, OPEX, and an implementation roadmap. Human oversight and operational control are explicit design requirements. 

Why it matters: this is what serious AI implementation in a safety-critical system looks like.

Rather than beginning with an AI model and looking for somewhere to install it, the project begins with:

functions → levels of assistance → human roles → ConOps → architecture → safety/risk analysis → technical specification → classification approval.

That is classic systems engineering applied to AI.

The approach is directly transferable to naval autonomy. A future USV, UUV mothership, or highly automated surface combatant needs precisely this decomposition. The important requirement is not “how much AI can we put aboard?” It is which functions can be delegated, under what conditions, with what authority, what fallback modes, and what evidence.

4. India’s GRSE couples defense shipbuilding AI with classification and open innovation

On 2 September, Indian defense shipbuilder Garden Reach Shipbuilders & Engineers signed an MoU with DNV covering advanced warships and specialized vessels, digital shipbuilding, AI, Industry 4.0, green technologies, and workforce development. The combination pairs GRSE’s warship design and construction capabilities with DNV’s classification, certification, technical-assurance, and digitalization expertise. 

That announcement coincides with a particularly interesting implementation mechanism. GRSE’s GAINS 2026 open challenge has “Artificial Intelligence in shipbuilding” as one of only three principal technology areas. Its first shipyard familiarization visits took place 3–4 September, giving outside innovators direct exposure to actual yard processes before they submit proposals. Selected projects can progress from concepts through detailed project reports and ultimately negotiations for implementation contracts. 

Why it matters: the combination matters more than either initiative alone.

GRSE is simultaneously building:

• an external innovation pipeline.
• direct exposure of AI developers to production problems.
• a potential pathway into funded implementation; and
• a classification/technical-assurance relationship surrounding digital shipbuilding.

That is a useful distributed-innovation model for naval shipbuilding. It reduces the temptation to create an isolated corporate “AI laboratory” disconnected from welders, planners, naval architects, production engineers, inspectors, and shipyard bottlenecks.

It also acknowledges that shipbuilding AI needs both experimentation and institutional assurance. Rapid innovation discovers useful applications; classification, engineering governance, and production qualification make them usable.

5. Caterpillar and FieldAI target the shipyard-like problem: physical AI in messy industrial environments

Caterpillar announced a collaboration with FieldAI on 2 September aimed at physical AI, autonomy, robotics, and facility digital twins. The collaboration combines Caterpillar’s engineering and operational data with FieldAI’s robot foundation models. Initial applications include autonomous inspection, facility and jobsite digital twins, situational awareness, and optimization through simulation and AI. 

The most revealing part of Caterpillar’s announcement is its emphasis on complex, dynamic industrial environments where traditional automation often falls short. FieldAI’s architecture is robot-agnostic and combines NVIDIA accelerated computing with Omniverse-based high-fidelity digital twins. 

Why it matters for shipbuilding: shipyards look much more like construction and mining sites than automobile assembly plants.

They contain:

• changing geometry.
• very large workpieces.
• people working alongside machinery.
• temporary structures.
• incomplete assemblies.
• variable lighting and accessibility.
• low production volumes; and
• frequent differences between nominal CAD and physical reality.

Conventional robotics thrives on repeatability. Naval shipbuilding frequently provides the opposite.

Caterpillar/FieldAI is therefore worth watching as a leading indicator for adaptive shipyard robotics. Autonomous inspection is the obvious first application, followed by material movement, surface preparation, machine tending, fitting assistance, and eventually selected welding and outfitting functions.

6. Hexagon brings AI into precision metrology—the missing half of AI CAM

Hexagon announced its next-generation OPTIV S optical CMM on 2 September. The platform combines higher machine dynamics and improved optical measurement with a new group of AI-assisted capabilities inside PC-DMIS. Hexagon reports approximately 30% higher machine dynamics and about 15% shorter inspection cycles in initial application tests. 

Of greater interest, AI Edge Detection uses AI-assisted camera/light calibration to improve measurement consistency and reduce programming effort and is scheduled for commercial release with PC-DMIS 2026.2. Fast Auto-Focus and AI Illumination are to follow in 2027. The new OPTIV S becomes commercially available on 14 September at IMTS and AMB. 

Why it matters: most discussion about AI CAD/CAM concentrates on getting from CAD to the machine. The more important industrial loop is:

CAD/MBD → CAM → machine → metrology → deviation → corrective action.

If AI generates a toolpath twice as quickly but inspection remains labor-intensive—or if dimensional evidence cannot feed back into the digital thread—the production system is only partly transformed.

For aerospace and defense components, AI-assisted metrology can reduce dependence on scarce specialists while producing the high-confidence measurement data that downstream AI systems require. In the longer term, the powerful combination is an AI CAM agent and AI-assisted CMM sharing the same PMI and persistent feature identity.

7. Aize’s acquisition of Samp tackles the “as-designed versus as-built” digital-twin problem

Industrial digital-twin company Aize announced on 2 September that it acquired French industrial-AI company Samp; it did not disclose the financial terms. Samp’s Shared Reality technology uses AI to turn laser scans and other reality-capture data into structured asset information and correlate physical objects with engineering tags, drawings, and P&IDs. Aize says Samp technology is already deployed at more than 500 industrial facilities. 

The strategic fit is unusually clear. Aize starts with rich CAD and engineering data for newer facilities; Samp starts with what exists in the physical plant. The combined product is intended to reconcile engineering intent, operational data, and field reality across greenfield and brownfield assets. 

Why it matters: this addresses one of the central weaknesses in most digital twins.

An authoritative model may tell you what the ship or facility was supposed to be.

A point cloud tells you what is physically there.

Maintenance records tell you what people believe has changed.

A useful lifecycle twin must reconcile all three.

That is especially important for ships after years of alterations, repair availabilities, temporary modifications, component substitutions, and battle-damage repairs. An AI system that compares reality capture against the engineering baseline could identify undocumented configuration differences and maintain a more credible as-maintained digital ship.

8. Research shows how factory AI may fit onto edge and air-gapped hardware

A 2 September research preprint addresses a less visible but important deployment problem: industrial AI assistants often need to run locally, while useful models can exceed the computational envelope of shop-floor hardware.

Researchers developed a method for selecting compressed sub-networks for retrieval-augmented factory agents based not merely on model size but on measured answer quality, device throughput, memory constraints, and minimum capability. In a manufacturing-manual experiment, retrieval-grounded distillation brought the compressed model back to within 4.6% of the unpruned model’s judged answer quality, and the system operated across three edge-device tiers at a reported 1.3–5 watts standby. 

Why it matters: many defense applications cannot assume continuous cloud access.

Think of:

• classified engineering spaces.
• naval shipyards.
• deployed ships.
• expeditionary repair facilities.
• air-gapped manufacturing systems; and
• Supplier environments with tight intellectual-property restrictions.

A locally hosted agent grounded in approved machine manuals, procedures, work instructions, and technical data may be far more useful operationally than a substantially larger cloud model it cannot securely access.

The architectural lesson is important: use the smallest model that performs the bounded industrial task adequately, ground it in authoritative information, and measure it on the hardware where it will actually operate.

9. Early market signal: AI is moving into physics itself

One more development is worth keeping on the radar. On 1 September, startup Physical Superintelligence launched with $58 million in seed funding led by Breakthrough Energy Ventures. Its proposed Emmy platform consists of AI-based “virtual physicists” coupled to curated simulations, initially aimed at optimization of terrestrial and orbital data centers. 

This is far earlier than the other developments in this briefing, and claims about designing systems beyond conventional engineering capability remain to be demonstrated.

Why it matters nonetheless: investment is moving from LLM-based engineering assistance toward physics-grounded AI. That is a potentially more consequential path for generative engineering.

The eventual prize is not a model that knows how engineers talk about heat transfer, fluid flow, structures, or electromagnetics. It is an AI capable of repeatedly invoking validated physics, exploring enormous design spaces, recognizing physical relationships, and returning candidate designs with the analytical evidence engineers need to evaluate them.

This is especially true for aircraft, ships, propulsion systems, antennas, heat exchangers, and autonomous vehicles, where generative design becomes genuinely transformative.

Across these developments, the recurring pattern is a shift from isolated AI capability to governed lifecycle integration. Product definition and naval-architecture tools establish the authoritative model; systems engineering and classification define permissible authority; robotics, metrology, and reality capture connect the model to production and the as-built asset; and edge deployment and physics-grounded AI extend that architecture into constrained environments and future design exploration.

Maritime and defense assessment

SMM made this an unusually maritime-heavy week, but the announcements reveal a coherent architecture rather than a collection of unrelated AI demonstrations.

NAPA is inserting AI into naval-architecture work. CMA CGM/SDARI/Bureau Veritas are wrapping AI-enabled ship functions inside ConOps, systems engineering, human factors, risk assessment, and classification. GRSE is simultaneously opening shipyard problems to outside AI innovators and developing digital-shipbuilding expertise with DNV. Aize/Samp tackles the as-built digital twin. Caterpillar/FieldAI tackles adaptive physical execution. Hexagon tackles measurement. And Autodesk is increasingly connecting the AI agent to the authoritative product definition.

Taken together, they suggest the architecture of a future AI-enabled ship enterprise:

Requirement → MBSE/ConOps → ship-design model → class/rules → PMI/MBD → production planning → CAM/robotics → automated metrology → as-built digital twin → operations → maintenance/reality capture → revised design baseline.

AI can act at almost every stage. But the digital thread, engineering rules, physics, verification, and configuration authority should remain the rails it runs on.

For naval shipbuilding, the near-term priority is to build the governed digital foundation before pursuing broad autonomy: authoritative product data, persistent configuration identity, model-based definitions, interoperable interfaces, validated engineering tools, and inspection evidence that closes the loop with the physical asset. Investment should favor bounded use cases that operate within this architecture, demonstrate measurable production or assurance value, and can scale across design, shipyard, supplier, and fleet environments. Assess upcoming demonstrations at IMTS and ICCAS against those criteria.

Main image caption: Vincent Malpaya, mechanical engineer with the Airborne Telemetry Group, compares a 3D-printed prototype to its computer-aided design (CAD) model at the Innovation Lab. (U.S. Navy photo by Ryan Figueroa)