When Robots Read Handwritten Instructions: FANUC Brings…

A handwritten note tells a robot which parts are needed. Cameras identify the objects, an AI agent interprets the request and an industrial arm prepares the kit. FANUC plans to demonstrate that chain at IMTS 2026 with Google Cloud, alongside a system that generates Python code and robot programs from natural-language commands.

The scenario sounds like a small convenience feature. In reality, it points toward a new interface for factory automation. Traditional industrial robots execute carefully engineered sequences. Physical AI adds perception and reasoning so that a system can translate an uncertain human request into a physical plan. The opportunity is faster deployment. The risk is that ambiguity leaves the screen and starts moving heavy machinery.

What FANUC announced for IMTS

FANUC America’s September 3 preview describes a broad set of demonstrations for the International Manufacturing Technology Show in Chicago, which runs from September 14 to 19. One application uses AI agents to interpret handwritten instructions, identify and locate components, and direct robots to build a manufacturing kit. Another, called CRX Vibe Coding, uses generative AI and spoken commands to produce Python code and robot programs.

Other demonstrations combine vision and force sensing for connector insertion, track bolts on moving parts, optimize machining paths through digital twins and attempt autonomous recovery after failures. FANUC names Google Cloud, NVIDIA, Amazon Web Services and Inbolt among the technology partners supporting its Physical AI work.

Why handwriting is a serious robotics test

Handwriting introduces uncertainty before the robot has moved. Letters can be unclear, part names abbreviated and quantities omitted. A human colleague uses context to ask a question or notice that an instruction makes no sense. A robot system must convert the note into structured intent, match that intent to an approved parts database and decide whether confidence is high enough to continue.

This makes the demonstration more than optical character recognition. The AI agent needs a controlled path from language to action. It should show the operator what it understood, identify missing information and request confirmation when the instruction falls outside an authorized template. The safest answer to an ambiguous note may be to do nothing.

Natural language is not the safety controller

A language model can help create a task description or draft code. It should not replace deterministic motion control, speed limits, collision monitoring and safety-rated hardware. Factories need a separation between the flexible layer that proposes an action and the verified layer that executes it.

That separation can resemble a compiler. The AI translates a request into a limited task representation. Software checks available tools, payload, reach, fixtures and forbidden zones. A simulator tests the trajectory. Only an approved program reaches the robot controller. If the environment differs from the validated state, the machine should stop or return to a known-safe routine.

Why established industrial robots matter

Physical AI is often illustrated with humanoids, yet most automated factories already use arms, machine tools and mobile platforms optimized for specific work. Adding better perception and easier programming to this installed base could create value sooner than replacing it with a general-purpose body.

An industrial arm already offers repeatability, payload and service support. The difficult part is engineering every new application: fixtures, vision, paths, error handling and connections to production systems. If AI can reduce that integration effort without weakening control, the economic impact could be substantial, especially for smaller manufacturers that cannot maintain a large robotics engineering team.

The digital twin becomes a gatekeeper

FANUC will also emphasize virtual commissioning and digital twins. A digital twin is useful here not because a virtual factory looks impressive, but because it can become a validation environment. Proposed paths can be checked for joint limits, collisions, reach and cycle time before the physical cell moves.

Simulation cannot prove everything. Real cables bend differently, parts arrive out of tolerance, surfaces reflect light and grippers wear. A trustworthy process therefore connects simulation with real sensor data and controlled commissioning. The twin narrows the risk; final evidence still comes from the machine and the actual workpiece.

From one programmed sequence to a library of skills

Conventional robot projects often produce code tied to one cell. Physical AI could encourage a higher-level library: locate a connector, align it, insert with force limits, verify seating and recover after a failed attempt. An operator would compose approved skills rather than define every motion point.

This approach can improve reuse, but only if each skill has explicit boundaries. The system must know which grippers, objects, tolerances and sensor conditions were validated. A label such as “insert connector” is not a universal capability when connector shapes, cable stiffness and access angles vary.

The data question moves inside the factory

Multimodal systems may process camera feeds, sensor values, handwritten notes and production records. That creates valuable training and diagnostic data, but also exposes drawings, part numbers, worker behavior and process knowledge. Manufacturers will ask where inference occurs, what leaves the site and whether their examples can improve models used by competitors.

Access controls and audit logs are therefore part of the robotics product. A plant should be able to trace the original instruction, the AI interpretation, the generated task, each approval and the final robot execution. Without that record, investigating a defect becomes much harder.

What would count as convincing evidence at IMTS

A live demonstration should include variation rather than one rehearsed input. Visitors should see different handwriting, a missing part, an ambiguous request and a deliberately unsafe instruction. The system’s refusals matter as much as its successes. It should explain what requires confirmation without producing a confident but incorrect action.

Useful performance measures include task-interpretation accuracy, operator corrections, programming time, cycle time, recovery rate and false-positive safety stops. FANUC’s announcement provides the concept and partners, but not independent results for these metrics. The exhibition is the beginning of the evaluation, not its conclusion.

Brownfield deployment will decide the market

Most manufacturers do not operate an empty, perfectly standardized facility. They have equipment of different ages, proprietary controllers and processes that cannot stop for a long AI experiment. Physical AI must work with this “brownfield” reality.

Successful products will integrate gradually. They may begin by generating programs for review, assisting vision setup or diagnosing a failed cycle. Autonomy can expand only after the system accumulates evidence. This staged route is less spectacular than a robot that understands every instruction immediately, but it aligns with industrial risk and maintenance practice.

The Alpha Bionic view

FANUC’s IMTS preview is important because it moves the Physical AI discussion toward machines that already produce goods. The headline feature is natural interaction. The deeper change is a possible new programming layer that sits between human intent and verified automation.

The winners will not be the systems that generate the most robot code. They will be the ones that make every generated action inspectable, reversible and limited by an approved safety envelope. In a factory, intelligence becomes valuable only when operators can trust both what the robot does and why it was allowed to do it.