A self-perform group runs a robotics pilot. In the vendor demo, a layout robot prints a full level of partition lines from the model in an afternoon, dead-on, and the room goes quiet in the good way. Six weeks later the same robot stalls on a real deck: the slab is out of level by more than the demo floor ever was, pallets sit on three of the control points, the model the robot reads predates last week's ASI, and the layout crew it was supposed to free up is watching it re-localize. The robot did exactly what it was built to do; the deck did exactly what real decks do. The gap was not mechanical but perception and decision: nobody had given the robot a reliable picture of where it was, what had changed, and what to do when reality and the model disagreed. That gap is the subject of this lesson. By the end you will produce the integrated robotics-plus-AI plan for a self-perform group: which robots, on which scopes, fed by which AI perception and decision layer, gated by which verification, tied to the prefab decision and productivity case you already own. The thesis is one sentence: AI is the perception and decision layer that makes robotics and prefab actually work together on a real deck, and without it you have an expensive demo.

The Pilot That Stalled on the Real Deck

The demo failed nowhere because the demo had no deck: a flat floor, a clean model, no other trades, no schedule pressure, a vendor engineer babysitting the machine. The real deck has none of those: it is out of level, congested, mid-revision, and shared with five other crews who do not care that your robot needs its control points clear. The mechanical capability the demo showed off was never in doubt. A layout robot marks points tighter than a human with a tape; a Boston Dynamics Spot can walk a deck and carry a payload; a robotic welder lays a bead more consistently than a tired ironworker at hour nine. The hardware works. What the demo hid is everything around it: knowing where the robot is, what the current model says, what changed, and what to do when the deck does not match the plan.

That surrounding intelligence is the perception and decision layer, the expensive, invisible part. A leader who buys robots and skips it has bought demo machines that perform in controlled conditions and stall in real ones, exactly the way the layout robot stalled when the slab was out of level and the control points were buried. The robotics question is not really a hardware question; the hardware is mature enough. The question is whether you can feed it a reliable, current, verified picture of the deck and design, and whether you have decided in advance what happens when the picture and the deck disagree. That is an AI and process question, not a robot-shopping question, and a plan that treats it as robot shopping will reproduce the stalled pilot at portfolio scale.

AI Is the Perception and Decision Layer

Here is the controlling analogy: a robot on the deck is a strong, tireless apprentice who is also completely blind and has no memory. It does exactly what you point it at, all day, to a tolerance you cannot match by hand, but it cannot see the deck, does not know the slab is out of level, the model changed last Tuesday, or a pallet sits on the point it is about to mark. Everything it cannot do, eyes, memory, and judgment, comes from somewhere else: the AI layer. Computer vision gives it eyes (where am I, what changed). The federated model and the CDE give it memory (the current design and latest revision). The decision logic plus the human gate give it judgment (proceed, flag, or stop when reality and the model disagree).

This maps onto the four engines. Computer vision reads the deck (OpenSpace and Disperse capture progress, Versatile and CraneView read the crane, a Spot walk turns a deck into a point cloud); predictive ML reads the trend (the data that flags schedule slip tells a robot whether the area is ready); generative design and the model supply the intent the robot executes; generative AI composes the reports and exceptions. Strip the AI out and you have four blind apprentices bumping into each other; add it, and the prefab shop, layout robot, Spot, and crane share one current picture of the job, the value of convergence, made of AI, not robots.

A robot on the deck is a strong, tireless, blind apprentice with no memory. The hardware is the muscle; AI is the eyes, the memory, and the judgment. Buy the muscle without the eyes and you have bought a demo, because the deck is out of level, mid-revision, and shared with five other trades, and the robot cannot see any of it on its own.

The Four Robotics Scopes a Self-Perform Group Can Run Now

A self-perform group does not adopt "robotics" but specific robots on specific scopes, each with its own readiness and AI dependency. The first is progress capture with Boston Dynamics Spot: a legged robot walks a programmed route on a schedule, carrying a 360 camera and sometimes a LiDAR payload, and feeds the capture into OpenSpace or Disperse for the same baseline-drift detection the field-operations lessons covered. The AI dependency is pure perception: the robot is a camera mount with legs, and the value is in the vision pipeline that turns its walk into a verified record. This is the lowest-risk scope because the robot never touches permanent work; a wrong capture is a bad report, not a bad weld.

The second is layout robots: a robot drives a total station and marks partition lines, anchor points, and penetrations from the coordinated model, replacing hours of manual layout. The AI dependency is perception plus intent: it must localize accurately on a real, out-of-level slab and read the current model, so the scope lives or dies on model currency and control-point discipline. The third is modular and prefab fabrication: the shop builds assemblies (MEP racks, bathroom pods, exterior panels) off the federated model, increasingly with robotic cutting, drilling, and assembly cells. The AI dependency is the model-to-fabrication handoff and QA: the assembly is only as good as the model, and a model error becomes a shop error then a misfit. The fourth is robotic welding with AI QA: a welding cell lays consistent beads, and a vision system inspects each against the weld procedure, flagging beads for a human inspector. The AI dependency is the path (joint geometry from the model) and the inspection (computer vision reading bead quality). In every scope the robot supplies the muscle and a named AI capability supplies the eyes, memory, or judgment.

AI-Driven Crane Control and the Versatile and CraneView Layer

The crane is the heartbeat of a vertical self-perform job, and the scope where the AI layer pays off without bolting on a new robot. Versatile and CraneView instrument the crane with sensors, a load cell on the hook, a camera on the boom, position and cycle data, and the AI reads the stream to tell you what the crane did: how many picks, what they weighed, how long each cycle took, how much of the day the hook was loaded versus idle, and where the bottleneck is. This is computer vision and predictive ML on the most expensive resource on the site, converting crane time from gut feel into a measured number. The 2026 productivity tooling extends this from after-the-fact reporting toward live cycle optimization, surfacing the rigging or staging change that shaves a cycle.

Crane control makes a specific point about autonomy and the gate. The crane lift is a life-safety operation: a pick over an occupied deck, a tandem lift, a pick in wind near the chart limit are where OSHA 1926 and the lift plan govern, and where a wrong call hurts people. So the AI layer reads and advises; it does not lift. Versatile and CraneView make the operator and lift director better informed, faster, and more measured; they do not replace the operator's hands or the lift director's authority, and a plan that quietly assumes they will has walked through the life-safety gate without noticing. The rule the program has carried since the cardinal-rule lesson applies in full: the more autonomous the action and the higher the stakes, the more verification and oversight scale with it, not down. AI-driven crane data is a productivity multiplier precisely because it stays on the read-and-advise side of that line.

The Convergence Compounds the Prefab Decision

Robotics does not sit beside prefab; it changes the prefab math, which is why this lesson ties back to the L3 prefab-versus-site decision and the L4 self-perform productivity case. The L3 lesson framed the choice as a trade-off: prefab moves work into a controlled shop, faster, safer, and higher-quality, at the cost of design lead time, transport, and the risk that a model error becomes a fabrication error. Robotics pushes every term. Robotic cutting and assembly cells make the shop faster and more repeatable, strengthening the prefab case, but they raise the cost of a model error, because a robotic cell will reproduce a wrong dimension across a hundred assemblies before anyone catches it. The convergence makes the model the single most consequential input and the AI QA on fabrication output non-optional.

For the self-perform leader, this reframes the L4 productivity case, which measured productivity in labor hours, cycle days, SPI and CPI, and the hours-saved-by-role delta. The convergence adds a lever (prefab content) and a dependency (model maturity). A group that can fabricate robotically off a high-LOD federated model and verify with AI QA can move more scope into the shop, off the critical-path deck into a parallel, controlled, measurable environment, a real gain in the same KPIs the L4 case used. But it is contingent: it only works if the model is mature enough to fabricate from and the verification rigorous enough to catch the error before the robotic cell multiplies it. So the integrated plan treats robotics, prefab, and productivity not as three programs but one decision: how much scope can we move into a robotically-augmented shop, fed by a model we trust and verified by AI QA we have validated, and what does that do to the firm's productivity numbers.

Autonomy on the Jobsite and the Life-Safety Gate

Every robot on the deck is an autonomous actor in a shared, hazardous space, so the convergence touches the life-safety gate directly and verification scales with autonomy. The progress-capture Spot is the easy case: it does no permanent work, so the safety question is mostly its movement (programmed routes, geofencing, stop conditions) and verification concerns its output. The layout robot is a step up: it marks the work the trades will build to, so a localization error or stale model propagates into mislocated work, and verification is the points it marked against the current model. The robotic welder and fabrication cell are higher still: they make permanent, load-path or life-safety work, so their AI QA is a life-safety verification, and a bad weld the QA missed is what the gate prevents.

The crane is the top of the autonomy-and-stakes ladder, and the rule there sets the rule for the whole plan: read-and-advise, never act, when life-safety is in play, human authority retained and verification scaled up, not down. This is the gate-not-mood discipline since L1, now applied to machines that move on their own. The integrated plan therefore carries an explicit autonomy-and-verification map: for each scope, the autonomy level (captures, marks, fabricates, advises), the life-safety exposure (none, propagated, direct, direct-and-immediate), and the verification and oversight that scales with it, from a Spot walk's output verification to a welding cell's life-safety QA to a crane that stays human-commanded with AI advising. Writing this down is how a robotics program stays a productivity story rather than turning into an incident report.

The Applied Problem: The Integrated Robotics-Plus-AI Plan for a Self-Perform Group

Here is the deliverable. Produce the integrated robotics-plus-AI plan for your self-perform group. Start with the scope inventory: for each of the four scopes (progress capture with Spot, layout robots, modular and robotic prefab fabrication, robotic welding with AI QA, and the crane layer via Versatile and CraneView), state whether you adopt now, pilot, or defer, and why, in terms of your actual trades and model maturity. Do not adopt "robotics"; adopt named robots on named scopes, each justified against the work your crews do.

Then write the perception-and-decision layer for each adopted scope: what AI capability feeds the robot (which vision pipeline, which model and CDE, which decision logic), how it gets a current and verified picture of the deck and design, and what happens when reality and the model disagree (the proceed, flag, or stop logic). This is the part the stalled pilot skipped, and what makes the plan survive a real deck. Tie it to the L3 prefab decision and the L4 productivity case: how much scope the convergence lets you move into a robotically-augmented shop, what that requires of the model, and what it does to the labor-hour, cycle-day, and SPI/CPI numbers the firm tracks. Then write the autonomy-and-verification map: for each scope, the autonomy level, the life-safety exposure, and the verification and oversight that scales with it, with the crane read-and-advise on the right side of the gate.

Close with the operationalization gate, the trap the running-pilots lesson warned about: a pilot that impresses in the demo and dies at scale. Name the go/no-go criteria that move a scope from pilot to standard practice (the productivity delta it must hit, the verification reliability it must demonstrate, the model-maturity threshold) and who owns the perception layer when the vendor engineer goes home. The deliverable is the integrated plan: scope inventory, perception-and-decision layer per scope, prefab-and-productivity tie-in, autonomy-and-verification map, and operationalization gate. The lasting product is a self-perform group that buys the muscle and builds the eyes, so its robots perform on the real deck, not just in the demo, because the AI layer gives them the verified picture the stalled pilot never had.

Key Takeaways

• The robotics question is not a hardware question. The hardware (Boston Dynamics Spot, layout robots, robotic welding and fabrication cells, instrumented cranes) is mature enough to work; the pilot stalls on the real deck because nobody built the perception and decision layer around it, the eyes, memory, and judgment the demo hid.
• AI is that layer. Computer vision gives the robot eyes (where am I, what changed), the federated model and CDE give it memory (the current design), and the decision logic plus the human gate give it judgment (proceed, flag, or stop). Strip the AI out and you have blind, memoryless apprentices; add it and the shop, layout robot, Spot, and crane share one picture.
• A self-perform group adopts named robots on named scopes, not "robotics" in the abstract: progress capture with Spot (lowest risk, pure perception, no permanent work), layout robots (perception plus model currency), modular and robotic prefab fabrication (model-to-fabrication handoff plus QA), and robotic welding with AI QA (path plus life-safety inspection), each with its own readiness and AI dependency.
• AI-driven crane control via Versatile and CraneView instruments the most expensive resource on the site and turns crane time into a measured number, with 2026 tooling moving toward live cycle optimization, but the crane lift is a life-safety operation, so the AI layer reads and advises and never lifts: the operator's hands and the lift director's authority are retained.
• The convergence compounds the L3 prefab decision: robotic cells make the shop faster and more repeatable (strengthening the prefab case) while raising the cost of a model error (a wrong dimension reproduced across a hundred assemblies), making the model the single most consequential input and the AI QA on fabrication output non-optional.
• It reframes the L4 self-perform productivity case by adding prefab content as a lever and model maturity as a dependency: more scope moved into a robotically-augmented shop is hours moved off the critical-path deck into a parallel, controlled environment, a real gain in the same KPIs, contingent on a model mature enough to fabricate from and verification rigorous enough to catch the error first.
• Autonomy touches the life-safety gate, and verification scales with autonomy and stakes, not down: a Spot walk gets output verification, a layout robot model-currency and control-point verification, a welding cell life-safety QA with human inspection of flagged beads, and the crane stays human-commanded with AI advising. The plan writes this map down so no machine sleepwalks through.
• The deliverable is the integrated robotics-plus-AI plan: scope inventory (adopt, pilot, or defer, by named robot and scope), perception-and-decision layer per scope (with proceed-flag-stop logic), the prefab-and-productivity tie-in, the autonomy-and-verification map, and the operationalization gate with go/no-go criteria and named ownership.