Google DeepMind has published a series of reports worth flagging for anyone whose work touches technical documentation:

• “RT-2: New model translates vision and language into action” (July 2023) — RT-2 is a Vision-Language-Action (VLA) model that translates visual information and natural language directly into a robot’s physical action. Where conventional robots executed only mathematical commands like “move from point A to point B,” RT-2 has been trained on vast amounts of web text and image data, allowing it to interpret abstract instructions like “place the dinosaur next to the healthy snack” and act on them autonomously.
• “Shaping the future of advanced robotics” (January 2024) — This report introduces AutoRT, SARA-RT, and RT-Trajectory — evolutions of RT-2 — covering a system architecture in which robots perceive their environment autonomously and decide actions safely under what the team calls a “Robot Constitution.”
• Google DeepMind’s Gemini Robotics page — This shows what happens when Gemini, Google’s most powerful multimodal model, becomes a robot’s “brain.” The point is that robots no longer simply execute predetermined motions — they read large amounts of data (Reasoning), understand situations multimodally (Multimodal), and translate this into physical action (Action).

What these reports point to, taken together, is a quiet shift in how text functions. Natural-language manuals — text data — are starting to look less like an aid for human comprehension alone, and more like a “blueprint of action” that robotic systems can act on directly.

If everyday human language is on a path toward driving robotic behavior, what might that suggest for technical writers working in the traditional sense?

The rest of this piece explores that question. It uses the term “Robot Knowledge Architecture (RKA)” as a shorthand for an expanded conception of technical writing. RKA is not an industry-standard term — it has been coined here for convenience.

Technical Writing vs. Robot Knowledge Architecture?

In the framing used here, RKA describes a knowledge system modularized in machine-readable form, structured so that a robot can perceive its environment and execute tasks based on it.

If conventional technical writing arranges information linearly (“if the temperature is high, turn it off”), an RKA approach is closer to designing a data pipeline: [sensor data] → [logical threshold] → [state definition] → [physical action vector]. (“If the temperature is high, turn it off” is illustrative; even as ordinary technical writing, it is a sentence to avoid.)

Consider: “In an emergency, press the red button.”

From a robot’s perspective, what does this look like?

• In an emergency (situation): What is an emergency?
• The red button (target): What does it look like?
• Press (action): How is it pressed?

To make this readable to a robot, we can hypothetically use Markdown syntax. In Markdown, {: } defines metadata such as classes, attributes, and IDs. (Note: {: } is an extension that works only in specific environments like Kramdown or Jekyll. In production, standardized approaches such as HTML5 data-* attributes or JSON Schema may be more appropriate.)

“If [the device temperature exceeds 80°C]{: logic=“temp > 80” unit=“celsius”} causing an [emergency condition]{: .emergency-trigger}, [press]{: .action-push} the [emergency stop button]{: #ESTOP_01 .target-feature=“red,circular”} firmly with [at least 5 kgf of force]{: value=”>=5” unit=“kgf”} to halt operation immediately.”

• exceeds 80°C (trigger condition): {: logic=“temp > 80”}
• emergency condition (situation): {: .emergency-trigger}
• emergency stop button (target): {: #ESTOP_01 .target-feature=‘red,circular’}
• at least 5 kgf of force (physical magnitude): {: value=”>=5” unit=“kgf”}
• press firmly (action type): {: .action-push}

DITA, with its topic-based structural model, aligns reasonably well with this kind of layered, condition-driven information, and could be one viable approach for RKA-based writing.

A caveat is worth registering. The strength of RT-2 lies in the “flexible generalization” learned through LLM pre-training. The fully tagged sentence above marks up every situational element, and at that extreme it can stand in tension with the model’s generalization. In practice, a technical writer would not need to redefine commonsense knowledge a robot already has. Tagging would concentrate where probabilistic inference is unsafe — hard numerical thresholds (80°C, 5 kgf), regulated parameters (ISO safety provisions), legal constraints — and leave the rest as prose.

From there, the same source content gets rendered twice: once for the robot, once for the human.

1. The robot’s reading (Machine Perception Layer)

An RT-2-class robot (a VLA model) strips emotional or stylistic modifiers from the text and extracts only the parameters that determine action, constructing a dataset that translates them into action vectors. (See report 1 above.)

[Logical Translation] : The robot’s reasoning process

• Trigger (condition): logic: “temp > 80” (Sensor_ID: Temp_01)
• Status (situation): class: “emergency-trigger” (Mode: EMERGENCY_LEVEL_HIGH)
• Target_Object (identification): ID: “ESTOP_01”
• Search_Filter (attributes): {color: “red”, shape: “circular”, label: “Emergency Stop”}
• Action_Type: type: “push” (Mode: Contact_Push_Force_Control)
• Action_Parameter: value: “>=5”, unit: “kgf”
• Flag (final intent): Termination_Sequence_Active

[Action Vector Mapping] : The robot’s physical output

• Vector Sequence: [1, Δx, Δy, Δz, Δr, Δp, Δy, 150]

2. The human’s view (Human Display Layer)

On a human-facing surface (manual, web page, AR display, etc.), the rigid tags and attribute values disappear, and the content appears in a maximally readable form:

“If the device temperature exceeds 80°C, press the emergency stop button firmly with at least 5 kgf of force to halt operation immediately.”

What Skills Would RKA-based Technical Writing Require?

If conventional technical writing has been about producing instructions for human readers, RKA-based writing — in the form this thought experiment points to — would be closer to designing a knowledge system in which humans and robots interact with the same source content.

Capabilities a writer would benefit from preparing for, in that scenario, fall along the following lines:

This does not mean a technical writer must necessarily acquire all such skills directly. A writer continuing to author manuals in natural language while an LLM performs initial structuring (auto-tagging), with the writer verifying the result for accuracy, is a workable division of labor that would still demonstrate the value of the role.

Can Robots Write Manuals for Human Users?

The “knowledge cycle” entry in the comparison above hints at another possible shift in the writer’s role.

Recursive self-improvement is the term used when a robot improves its own performance based on a combination of pre-existing data and new data gained from acting in the world.

Borrowing the concept loosely, a robot that has completed simulation could surface edge cases (for example, “at this angle, there is a hand-pinch risk”) and produce a guide or real-time warning content for human operators based on what it has learned. The technical writer’s role in such a loop would shift toward verifying what the robot has drafted and releasing the final document.

Technical Writers, Designing Robot Intelligence

We no longer live in an era of paper-only manuals. As Google’s RT-2 case suggests, a technical writer’s text can become a robot’s experience, and a robot’s experience can in turn become a guide for us.

This article has examined how the technical writer’s role may be expanding from “manual production” toward “designing the information structure for robot intelligence.” RKA, in part, is an area in which traditional technical writers would need to acquire new technologies and tools. But programming engineers cannot fully replace this role — because the core capability is not simply writing or coding, but designing the structure of information itself.

Not every technical writer needs to transform into a designer of robot-intelligence information structure. But in an era where intelligent robots and humans coexist, and viewed through the lens of “Robot Knowledge Architecture,” the technical writer can be seen not as a profession in decline or stagnation but as one with a new and more expansive set of opportunities ahead.