The Physical AI Primer.

Four parts: concepts, robots, people, reference.

Each robot has its own section: the photo first, then the details.

Robot sections are numbered R01 to R22. People sections are C1 and C2. Use the contents panel and the robot index to jump to a section.

Ten facts that frame the robot industry.

The essentials a decision-maker needs before any conversation about robots on site.

People lead the work; robots handle defined tasks.

"Human in the lead, robots in the loop" is the operating model used throughout this primer.

AGC (2026): "Technology is unlikely to eliminate the industry's need for people." A Meta spokesperson, as reported in Aug 2026: "We need more workers, not fewer." OSHA notes many robot accidents occur in non-routine work such as setup and maintenance.

Eight ways people and robots share work.

A working taxonomy by who holds authority and where the person stands. Construction today sits in the middle rows.

In construction, most deployed systems are operator-directed or supervised autonomy. No signed US construction labor-management robotics agreement was found (as of Oct 2026).

Fourteen terms cover most of the robot talk.

Working definitions used throughout this primer. Autonomy is a spectrum, so most terms describe a type of machine, not a fixed ability.

Four forces are pulling robots into the field.

Labor, capital, AI models and hardware cost all moved at once since 2023.

Counterpoint: the lowest-priced humanoid listed (Unitree G1) is near $13,500 (single source), while Agility's Digit 5 bill of materials is about $150,000 (reported). Cost depends on class.

From one factory arm in 1961 to humanoids in 2026.

Only dated, verified milestones. Each step widened where robots can work.

Ten jobs cover what robots do on site today.

Grouped by the work performed, with typical examples and the maturity label for each group (as of Oct 2026).

Form factor decides where a robot can go.

Four families, from fixed arms to heavy equipment. Mobility sets the site preparation each one needs.

Rodney Brooks expects the "humanoid" label to persist even as designs drift toward wheels and varied grippers (Sept 2025).

Size sets the site rules: Light, Medium or Heavy.

Class follows the task and the site rules, not weight alone. Examples show published specs (vendor figures).

Okibo's 800 lb finishing robot is Medium; a 74.5 lb quadruped is Light. Brooks advises staying 10 ft (3 m) from a full-size walking robot (opinion, not a rule).

Autonomy is a ladder of five levels, not a switch.

The same robot can run at different levels by task. Level names are fixed across this primer.

1. Level 1: Teleoperated

   A remote human commands each move in real time. Also used to gather training data.

2. Level 2: Assisted

   Human leads; the machine guides or automates a step. Example: Hilti Jaibot drilling.

3. Level 3: Supervised autonomy

   Robot runs the task; a human approves it and handles exceptions. Examples: Built, Teleo, Spot.

4. Level 4: Autonomous in defined tasks

   No operator inside a set domain. Examples: warehouse AMRs; mining haul trucks (Cat Command, Komatsu FrontRunner).

5. Level 5: Fully autonomous

   Open-ended work with no human help. No example found in public sources (Oct 2026).

Rule used in this primer: label each deployment by the level it runs at today, not by the level a vendor targets.

Maturity differs sharply by robot family and site.

Primer assessment from deployments found in public sources. Darker means more mature use (as of Oct 2026).

Legend: None found; Pilot; Early comm. (shown as Early); Commercial (shown as Comm.).

No public data center deployment of humanoids found (as of Oct 2026). The one long-running data center reference is a single ANYmal inspection site. Mining haulage autonomy has hauled over 3 billion tonnes (Cat, 2021).

Only 6 of 20 featured robots rate Commercial.

Class and maturity for each robot profiled (as of Oct 2026); primer assessment. Where weight is not disclosed, class follows task and site rules.

China installed 54% of the world's robots in 2024.

Industrial robots dominate unit counts; service robots and humanoids are growing from a far smaller base.

Record funding sits beside visible financial stress.

Valuation is not funding raised: figures below are as reported, and several are unconfirmed or differ by source.

Real adoption is concentrated; hype is broader.

Evidence of use on one side, claims ahead of output on the other. Seven topics, same order.

Verdict: scale today sits in arms, warehouse mobile robots and service robots. Humanoids, construction robots and data center robots remain early.

Eight challenges decide how fast robots scale.

Each challenge with the strongest public evidence. Many figures are vendor claims or single sources.

No independent benchmark exists for humanoid reliability; most figures here are vendor claims or single sources.

2030 forecasts for humanoids differ by nearly 5x.

Annual humanoid shipments in 2030 under three analyst views. The spread reflects scope, cost and use-case assumptions.

Forecasts are analyst scenarios, not output. Goldman's 2024 report predates the 2025-26 funding wave.

Robot brains attract the biggest valuations.

Foundation models and compute for physical AI; mostly early access, little deployed proof.

Thor early users named by NVIDIA: Agility, Amazon Robotics, Boston Dynamics, Caterpillar, Figure. DeepMind says movement speed still needs improvement.

Billion-dollar rounds in 2026 go to a few firms.

Reported 2026 financings; deal status differs, so the bars are not like-for-like.

Announced milestones run from Q4 2026 to 2028.

Company or press targets, not delivered output.

Dogotix deliveries begin in 2027 per the company. Slip risk for each item is shown in the next section.

Slip risk is highest where volume is promised first.

Slip risk is a judgment based on the cited evidence, not a sourced probability.

Not rated: Gemini Robotics 2 broader availability has no date. Figure and NVIDIA investment talks are reported talks only.

Forecasts for 2030 span 250,000 to 890,000 units.

Humanoid forecasts differ by source and date; none names data centers or construction as an early volume segment.

Goldman's 2024 report puts unit cost at $30,000-$150,000, down from $50,000-$250,000 a year earlier. Its early segments are logistics, warehousing and automotive.

Forecasts need growth with no precedent.

Today's units are small, and reliability is below what factories ask for.

Forecasts assume volume arrives first where tasks are defined; reliability and unit cost decide how fast it spreads to sites.

What could speed adoption, and what could stall it.

Evidence gathered in this research; the balance is not forecast.

Small, adaptable machines for repetitive tasks have been the most successful on jobsites so far (Construction Dive, Aug 2026).

A robot is five layers plus a site system.

Read top to bottom: from what the site manager sees to the parts that move.

Actuators are the largest cost in a humanoid.

Estimates put actuators at 35-60% of bill of materials; sources differ. Gear and screw types trade precision, shock and cost.

Torque density by technology (Nm/kg, whole actuator) has no primary source. A target above 10 Nm/kg is cited for humanoid leg actuators (vendor blog).

Sensing splits into seeing, touching and locating.

Four jobs, four sensor families. Construction robots add survey-grade positioning on top.

Lidar price, radar and inertial-unit grades are not benchmarked in public sources. Ask vendors for tested range, accuracy and failure behavior.

Layout robots locate themselves against the BIM.

The pattern used by the Hilti Jaibot drilling robot and Dusty layout printers: one model, one surveyed frame.

1. Model

   Design data in Revit or AutoCAD (Dusty also reads IFC) is the single source of truth. (Design)

2. Upload

   The model goes to a cloud, and a job file goes to a field tablet. (Cloud)

3. Set frame

   A total station on surveyed control points fixes the site coordinates. (Survey)

4. Locate

   The robot measures its position against the total station, so against the model. (Robot)

5. Work and sync

   The robot drills or prints, then syncs status back for documentation. (Execute)

Most robots run 1.5 to 5 hours per charge.

Verified figures only. Runtimes are vendor claims under unstated duty cycles; real runtime under load is likely lower (analyst inference).

Swap versus fast charge changes site infrastructure: swap stations and spare packs on one side, charge docks and electrical load on the other. Both are on the market.

Hands are costly; simple grippers still compete.

Reported hand designs

Digit 5 swappable grippers, Jaibot drills and Dusty printers show the same pattern on sites: the tool is matched to the task, not to the human hand.

Fast control loops run on the robot itself.

NVIDIA Jetson Thor is the reference part. Figures are vendor specifications.

Robot foundation models: claims, not benchmarks.

All results below are company-reported. No standardized, independent robotics benchmark exists (noted by several commentators).

As of Oct 2026. Funding context: Skild raised about $1.4B at over $14B (Jan 2026); Figure raised over $1B at $39B (Sep 2025).

Simulation trains robots before the site sees them.

Simulation lowers the cost of trial and error, but results on site still depend on supervised trials with real data (see Section A2.14).

Hand-held grippers beat teleoperation by about 3x.

The UMI gripper collected 111 demos an hour against 35 for SpaceMouse teleoperation; bare hands reached 231.

UMI (Stanford, Columbia, TRI; RSS 2024) is a hand-held parallel-jaw gripper with a GoPro. Bare hands is the human reference rate; the study also reports policy transfer between UR5e and Franka arms.

Control stays onboard; the network is for fleets.

Owners with federal or critical-infrastructure exposure should treat robots from listed vendors as high risk. Orbit's product page does not address security.

Humanoid safety standards are still being written.

Draft ISO 25785-1 topics: fall zones, controlled descent on power loss, contact-force limits and zoning. Most current humanoid deployments are fenced or zoned (single source). Safety-rated performance levels: verify against ISO 13849 before quoting.

Demos and sites differ in what gets measured.

Targets slip as well: Tesla's January 2025 goal of about 10,000 robots was missed (Electrek), and Agility's $300M+ of orders is milestone-contingent.

Deployment runs from scoping to supervised scale.

Sequence synthesized from vendor and standards sources; no vendor publishes a single canonical plan.

1. Scope

   Pick tasks with bounded variation: carry, layout, inspection, tote moves.

2. Baseline

   Import BIM or CAD; scan the site to build or check a digital twin.

3. Site frame

   Set control points (total station, RTK outdoors) so robots and BIM share coordinates.

4. Assess safety

   Risk assessment to ISO 10218-2; fall and exclusion zones; e-stop.

5. Prepare site

   Charge docks or swap stations, network backhaul, storage and spares.

6. Pilot

   Simulate, run supervised trials, collect corrections and tune skills.

7. Run and scale

   Humans lead, robots in the loop; log interventions; update fleet.

Data centers: a few proven tasks, many pilots.

Status of robot work in data center construction and operations, as of Oct 2026 (author synthesis of cited evidence).

Meta has acknowledged that robots cannot yet handle the cabling intensity of Nvidia GB300 systems (reported, Aug 2026).

Task by task, most data center work is still manual.

Phase, task, robot type and maturity. Maturity is an author synthesis, not a vendor scale.

Cost and workforce figures conflict by source.

Published numbers for building one megawatt (MW) of capacity. They are shown side by side, not averaged.

Sources differ. No credible public figure was found for cable miles, racks, anchors or welds per MW.

Peak workers per MW differ by more than 10x.

Illustrative: peak or on-site workers divided by capacity. Inputs are sourced; scope differs by case.

Abilene uses current workers against full-operation capacity; Hyperion uses expected peak workers. Neither is a like-for-like peak.

The best-documented data center construction robot.

Status: late pilot. Drills rack-stop and hanger holes in concrete slabs. All performance figures below are vendor claims.

DEWALT results are vendor-only and partly inconsistent.

What was reported, and what the sources do not settle.

Layout, capture and piling are in field use.

Named construction-phase deployments with data center links.

The Hyperion building-site article lists no robots; Buildots AI progress tracking is the only named technology there.

Hyperion shows the scale robots must meet.

Richland Parish, Louisiana. Turner, DPR and Mortenson joint venture; sitework began in 2024.

Robots are named on the power supply side (solar piling), not yet on the data center building site.

Inspection patrols are the most mature operations use.

ANYbotics ANYmal at the Digital Realty Zurich campus, in use since July 2022.

Meta tests robots for rack-side work; limits remain.

Reported trials (WIRED reporting via secondary outlets, Aug 2026). Many claims rest on anonymous sources.

An anonymous worker estimated cable-swap robots could someday remove up to 80% of the workload in some roles. This is an employee estimate, not a Meta forecast.

Emerging players are early, with thin public proof.

Status and funding as of Oct 2026. Vendor claims are labelled.

Roze's reported ask is a $100B valuation with no disclosed product; its assets would include ABB Robotics ($5.4B deal, 2025) and SB Energy.

SoftBank redesigns the rack for robots.

The cableless robot-friendly rack announced Sept 2025 removes the cabling task instead of automating it.

1. Power

   Busbar power replaces power cords at the rear of the rack. (No cords)

2. Optics

   A blind-mate optical connector replaces patch cables. (Blind mate)

3. Cooling

   A blind-mate liquid connector joins the server to a manifold. (Blind mate)

4. Carry

   Container robots pass servers to lifter robots. (Handoff)

5. Seat

   Lifter robots seat the server using QR codes and lasers. (Position)

No cable-pulling robot is verified as of Oct 2026.

Searches of trade press and vendor pages found no commercial cable-pulling, tray or conduit robot.

Absence from public sources is a finding, not proof of absence. Welding vendors (Path, Hirebotics) show data center weldments but name no data center customer.

Repetitive, geometry-defined tasks lead.

Task, robots and maturity on non-data-center sites. Figures are vendor-reported unless noted.

Independent productivity studies were not found. Dusty's own Turner pilot was cost-neutral on one project.

GC pilots are real; results are vendor-reported.

Only deployments with a named contractor and a cited source are shown.

China installed 54% of the world's new robots.

New industrial robot installations in 2024, top six countries (units).

Robot density: Korea leads; the US ranks eighth.

Industrial robots per 10,000 manufacturing employees, 2024 data (single-source summary).

Summary of the April 2026 IFR density release; the primary press release was not retrieved.

Humanoid factory work is pilots and plans.

Status as of Sep-Oct 2026. One tracker counts 2 of 48 humanoid designs as deployed, 8 as piloting.

Goldman Sachs projects humanoids could fill about 4% of US manufacturing labor shortages by 2030 (secondary report).

BMW and Figure: best-documented humanoid factory case.

Two Figure 02 robots in the Spartanburg body shop. Two robots, one task: proof of concept, not of scale.

Hyundai's Atlas plan runs to 2030, still unverified.

Dated milestones are Hyundai plans or press reports, not deployments.

Reports of "25,000 Atlas units" (installed total) and "30,000 a year" (capacity) are different metrics; the 25,000 figure rests on one aggregator citing Korea Times.

Humanoid pilots target tasks that stay manual.

Industrial arms are the proven core; humanoids aim at work that is hard to fixture.

Tesla Optimus lines and most Chinese humanoid volumes are company targets or research sales, not verified factory work.

Amazon runs the largest robot fleet.

Robotics program began with the 2012 Kiva acquisition. Figures as of July 2025 unless noted.

NYT-reported internal documents suggest Amazon aims to automate 75% of operations by 2033; Amazon says they were idea exchanges, not settled strategy.

Warehouse robots are proven; humanoids are early.

Scale is vendor-reported unless noted.

IFR: logistics is 47% of professional service robot shipments (117,500 units in 2025).

Solar is the most advanced outdoor construction segment.

Robots install modules and piles at utility scale. Figures are company claims unless noted.

AES claims Maximo needs half the time and cost of conventional methods (unverified). Crew size and cost were not stated. AES said it did not change hiring practices.

Mining and ports automate; labor shapes ports.

Verified cases. Several are dated; current statistics were not found.

Maturity is highest where work is repetitive and mapped.

Maturity of robot types by sector, as of Oct 2026 (author synthesis of the evidence in this section).

Legend: None / n.a., Pilot, Early comm., Commercial.

None = no public deployment found as of Oct 2026. n/v = not verified in sources read. n/a = not applicable.

Robots work where tasks are repetitive and mapped.

Seven conclusions from the application areas, as of Oct 2026.

Few humanoids can be bought in the US today.

Humanoid suppliers by headquarters, route to purchase and US availability, as of Oct 2026.

Sources conflict on some Chinese pricing; ranges are shown in the price-point section.

Shipping, pilot, waitlist and not yet differ widely.

Where named suppliers sit on availability to US buyers, as of Oct 2026.

Only Chinese hardware shows immediate stock across categories, and it carries policy risk. Construction-specific vendors rarely publish price or lead time.

Site robots are sold by quote, service or dealer.

Main suppliers by category: headquarters, how to buy and US availability (Oct 2026).

No vendor read published a California service-level commitment; California coverage is the least documented column. Skydio drones are sales-gated with no public price.

Published lead times are rare and mostly vendor claims.

The few weeks or dates found in public sources, as of Oct 2026.

Trade rules shift every few months.

Events that affect robot cost and supply, in order (as of Oct 2026).

Trade status: some facts firm, many still open.

What is confirmed, what is not, and where components concentrate (as of Oct 2026).

A June 2026 Section 232 update listed no robotics action. Treat the robotics outcome as pending until a proclamation is confirmed.

Five ways to acquire a robot shift risk differently.

Who owns the machine, how payment works and who carries the risk, with published examples.

Rent and RaaS move residual value, obsolescence and maintenance to the vendor. Buying keeps them with the customer.

Own for steady use; pay per use for lumpy use.

Buy and lease versus rent, RaaS and outcome-based pricing.

Digit break-even: $107,000 a year ($8,500 x 12 plus $25,000 deployment over 5 years) divided by $26.26/hr loaded hand-laborer cost is about 4,075 hours a year, before supervision and downtime. Illustrative.

Hours per year and project length pick the model.

Illustrative guide for one robot, before supervision, downtime and tax. Not vendor or lender quotes.

Buy suits steady use if residual and upkeep risk are acceptable. RaaS beats national-wage labor only near two shifts; higher local wages lower the hours needed.

Prices span about 300x; many are not disclosed.

Reported and list prices in USD by class, as of 2026. Sources differ for the same model by 30% to 40%.

Installed cobot cost is reported at 2.5x to 4x hardware. Hardware is a minority of total deployed cost.

Jaibot is rented on terms that favor Hilti.

Key terms from Hilti US Jaibot System terms (read Oct 2026). Prices sit in each Order Form and are not public.

Reported rate: $10,000+ a month (single source, Buildermuse, Sep 2026). The terms contain no data-ownership clause.

Lease structure decides cost, tax and exit.

General framework; robot-specific rates, residuals and advance rates are not publicly verified.

Rate context: one July 2026 SMB-finance site lists fixed equipment rates of 6.25% to 6.45% on $100,000-$249,999 (single source, low reliability). Robot-specific residuals are not published.

Digit RaaS terms come from one reported source.

Reported terms and illustrative arithmetic. Source: Dynamo reading of the Churchill XI S-4 (single source).

Contracts move from pilot to fleet in stages.

Published examples of each stage; most SLA and data terms are not public.

Insurance is bespoke; warranties are short.

What is published about cover and warranty for robots, as of Oct 2026.

Other warranty terms, battery life and residual values were not found in public sources.

Federal expensing is clear; California differs.

Tax and incentive rules relevant to a robot purchase, as verified (Oct 2026). Not tax advice.

Practical effect for a California taxpayer: full federal year-one deduction, smaller state deduction. Depreciation class for robots was not confirmed.

Ten points to settle before signing.

A procurement checklist drawn from the terms and gaps found in public sources.

Check supplier health, support, safety, data and record.

Six scorecard criteria, and reported distress cases showing why they matter (as of Oct 2026).

Each example is one filing or report at a point in time, not a verdict. Digit 4 passed an OSHA-recognized lab evaluation; Digit 5 third-party certification for work outside fenced cells is pending.

A robot is only as ready as the site is.

Vendor workflows start with survey control, model data and a map before any robot works.

1. Survey

   Jaibot and Dusty depend on a surveyed total-station control network on the floor. (Hilti, Dusty)

2. Prepare data

   Dusty accepts RVT, DWG and IFC; Jaibot uses AutoCAD or Revit data sent through Hilti Cloud. (BIM / CAD)

3. Map

   Agility Arc handles facility mapping and workflows; Spot Orbit records missions and map routes. (Arc, Orbit)

4. Connect

   Orbit runs over Wi-Fi, LTE or cellular, or on an on-site rack. Site-wide Wi-Fi or private LTE is an inference. (Network)

5. Stage power

   Dock locations and spare batteries are site-planning items (inference). Digit docks on its own. (Charging)

Fleets need new roles and a software layer.

Human in the lead, robots in the loop: people set the work, triage exceptions and repair.

Verified integrations are few and vendor-described.

Only examples found on vendor or press pages are listed. Everything else is marked not verified.

Oxmaint figures (8 to 16 month payback, under-5-minute work orders) are marketing claims without cited sources. Dusty has no publicly documented API.

Arm service intervals depend on model and shifts.

Figures come from secondary summaries or one vendor presentation. The model manual governs.

Derived: 20,000 hours is about 10 years at 2,000 hours a year (one shift) and about 3.3 years at 6,000 hours a year.

Outside arms, public service data is thin.

Warranty terms, consumables and who may repair are documented more often than service intervals.

No verified service interval was found for humanoids, quadrupeds, layout robots or excavator autonomy kits.

Service is usually bundled, rented or a yearly fee.

Published plan prices are rare. The figures below are the ones with a source.

No verified uptime benchmark exists for most classes.

No verified MTBF or measured uptime was found for Digit, Spot, Jaibot or Dusty (as of Oct 2026).

Runtime is short; cycle life sets replacement.

Runtime figures are vendor specifications. Cycle-life years are simple division (derived).

Calendar aging and temperature also limit pack life and are not modelled here. Digit 4 runtime differs by source (up to 8 h in 2024; up to 4 h reported in 2025).

Physical life, support life and tax life differ.

The bars are different measures for arms, not one lifespan. Humanoid and legged life: none verified.

Controllers can age out before the steel does.

Resale evidence covers refurbished arms only. Software support horizons are the nearer limit.

Measure output per hour, not robot count.

Only the first three measures come from a vendor KPI set. The rest are suggested measures (inference).

Baseline the manual process first. Without a pre-robot cost per unit, a payback claim cannot be checked.

Rule of thumb: budget at least 3x the machine price.

BCG-cited rule: budget at least 3x the machine price, and 4 to 5x for larger overhauls.

Example from the BCG-cited rule: a $65,000 six-axis robot implies a budget of about $195,000 (3 x $65,000). Simple ROI ignores downtime and scrap or rework gains.

Labor rate and throughput drive both cases.

Illustrative. Inputs are sourced; throughput is a vendor claim; excludes consumables, setup, lifts and a second worker.

A: manual 200 holes per shift (150-250 range); 500 holes is the vendor claim; $10,000+ a month is a single-source report. At LA scale the break-even is 64% of the claim; at the US median, 98%. B: 20-day month assumed from Hilti's Jaibot terms. *Reported results are far smaller: 50-70% less layout labor; cost-neutral at Turner.

Humanoid and quadruped cases need heavy use.

Illustrative. Assumes one robot-hour equals one human-hour, which is unverified and the largest uncertainty.

C: human rate uses BLS hand laborer median $18.38 and the 70.0% wage share of compensation. Digit prices are Dynamo assumptions from one source; hardware $200,000 is not a filing figure. D: single secondary source, unverified.

Published paybacks are mostly vendor claims.

None below is an independent audit. Each is labeled by who made the claim.

Jaibot savings erode below about 5,000 holes and pencil out above about 10,000 (single source). Data center ceilings may need 15,000 to 50,000 penetrations.

Reported ROI misses cluster around site, volume, cost.

Reported causes, with the evidence behind each.

Decide in four gates before any rollout.

A yes to every item in a column clears that gate. Open items become contract terms.

Supervision ratios beyond one operator per robot are earned from measured intervention rates, not assumed (inference).

Humanoid safety standards are still being written.

Status of the main robot safety and security standards, as of Oct 2026.

Four modes set how robots and people share space.

ISO 10218-2:2025 now holds the collaborative requirements that were in ISO/TS 15066.

Under the 2025 edition a robot is not "collaborative" on its own. Only a specific application (robot, tool, workpiece and task) can be designed, tested and confirmed as collaborative.

Operating models run from caged cells to fleets.

Who holds authority and where the human stands. A working taxonomy, not a standard.

In construction, most deployed systems today are operator-directed or supervised autonomy, not shared-task collaboration.

A site safety case is built in six steps.

Typical sequence for a robot on a construction or data center site. General practice, not a legal checklist.

1. Define

   Specify the tasks, the site and the operating environment, including floors, slopes and zones. (Scope)

2. Assess

   Run a task-based risk assessment under ISO 12100 and RIA TR R15.306 across the robot lifecycle. (Hazards)

3. Reduce

   Design out hazards, then add guarding, zones, fall areas and e-stop and power-loss behavior. (Controls)

4. Validate

   Test stability, stops and battery hazards; seek third-party inspection where available. (Proof)

5. Authorize

   Train staff, integrate lockout/tagout and set user responsibilities (R15.08-3, TR R15.706). (People)

6. Operate

   Log incidents, re-validate after changes and review the case on a fixed cycle. (Ongoing)

OSHA has no robot rule; general duties still apply.

Only verified points shown; items marked general framework are background, not legal advice.

EU rules arrive in 2027; China set a standards map.

Two regimes a global supplier must plan for. Dates as of Oct 2026.

Both regimes are moving: EU duties start in 2027-2028, and China's standards were still mostly draft in early 2026.

Liability follows the role each party plays.

General framework for US deployments; not legal advice. Insurance market data for humanoids was not found.

Robot-as-a-service is described by IFR as already the dominant business model in the US, so who counts as "operator" is a contract question to settle early.

Incidents and suits are mostly allegations so far.

Verified reports only; claims are allegations unless a finding is stated. Outcomes not found as of Oct 2026.

Baseline: NIOSH counted 41 US robot-related workplace deaths from 1992 to 2017, mostly industrial arms. No humanoid recall was found.

Connected robots add cyber and privacy exposure.

Security findings are researcher reports; privacy rules are a general US framework, not legal advice.

Key dates through 2028 shape the rules ahead.

Published and expected dates for standards and regulation, as of Oct 2026.

Open items: no U.S. robotics executive order was found, H.R. 8189 (foreign-made humanoid procurement ban) is only proposed, and the Section 232 robotics outcome was not found.

Check technology, site and safety before committing.

Decision checklist before deploying robots on a large project.

Then test commercial, workforce and supplier risk.

The second half of the decision checklist.

Ask each supplier for measured uptime, intervention rates, financial position, service coverage and a safety file. None of the vendors reviewed publish measured uptime.

Ten things to remember about robots.

The essentials from Part A, as of Oct 2026.

Robots R01 to R12: inspection, humanoids, layout.

Each robot starts with its photo.

Status as labeled on each robot section; see the maturity map in A1. As of Oct 2026.

Robots R13 to R22: finishing, printing, heavy machines.

Each robot starts with its photo.

Status as labeled on each robot section; see the maturity map in A1. As of Oct 2026.

Quadrupeds lead; only Digit has paid humanoid work.

Class, status, price signal, payload and runtime for each featured robot (as of Oct 2026).

Two quadrupeds have data center references; none of the five humanoids does (as of Oct 2026). ANYmal: one documented site, 2022 onward (2023 vendor case study). Spot: 2023 reports. Digit is the only humanoid in paid use.

Reported values range from $2.5B to $39B.

The three humanoid makers with a reported value; the bars use different bases, so read them as signals only.

Valuation, funding raised and announced deals are different measures. Apptronik has raised nearly $1B; Agility $390M+ before the SPAC; Figure about $1.8B to $2B depending on source.

ANYbotics ANYmal D

ANYmal is a rugged inspector for process plants: a Swiss quadruped with an IP67 body, built for scheduled rounds in industrial sites.

ANYbotics lists power, mining, oil and gas and chemicals as its industries; its one data center case study dates from 2023.

1. Map (Onboarding)

   An operator walks the site once; gauges, pumps and panels are tagged as inspection points.

2. Schedule (Shift Fleet)

   Missions start from a timetable, the ANYmal API or a control-system alarm.

3. Patrol (Autonomous)

   The robot walks the route, avoids obstacles and captures visual, thermal and acoustic data.

4. Analyze (Shift Insight)

   Shift Insight compares readings with operating data to flag heat, leaks and misalignment.

5. Act (Shift Connect)

   Confirmed anomalies become work orders in the customer maintenance system.

ANYmal runs a five-step inspection loop. The robot learns a route once, then repeats it and sends findings to maintenance systems.

Shift (launched Oct 1, 2026) replaces the earlier Data Navigator product and is planned to support third-party robots.

Price is on request; integration is the real work. Alternatives in the class: Boston Dynamics Spot, Ghost Robotics Vision 60, Unitree Go2 and B2. For fixed points, fixed sensors may do the same job.

Boston Dynamics Spot

Spot is the best-documented Western quadruped, sold to plants, utilities and builders; Boston Dynamics claims over 1,500 robots in customer hands (vendor claim, unverified).

Fleet size is not independently verified: Unitree shipped about 23,000 quadrupeds in 2025 alone. Boston Dynamics named Rohit Prasad CEO effective Oct 7, 2026; no Spot discontinuation has been announced.

Spot is proven at inspection, not at manipulation. Task maturity and named deployments.

Boston Dynamics no longer lists data centers on its Solutions page (Oct 2026). No 2025-26 named data center customer was found and the 2023 fleets are unverified. Vendor claim: over 90% less inspection time on construction sites.

Spot costs beyond the robot are not public. Economics, integration and labor effects (as of Oct 2026).

Current data center references are unverified: the Novva, Kio and Oracle Labs reports date from 2023.

Agility Robotics Digit

Digit is the humanoid with paid warehouse work. Digit 5 was unveiled Sept 15, 2026, with early access in 1H 2027 and general availability at end-2027.

2025 net sales were about $1.8M per Agility's filing (one trade headline misprints $18M). Digit 5 weight and specs are vendor figures, not independently verified.

Proven work is tote handling; the rest is planned. The $300M+ Digit 5 order book depends on contract milestones and is not revenue; most is reportedly one three-year contract for 1,000 robots.

Digit is priced to rent; purchase is an analyst case. Integration: charging station, a work cell for Digit 4, WMS/WES/MES links. Labor: augments tote moves; supervisors and technicians remain.

Limits: under 100 robots built a year; 2025 operating loss about $140M; the SPAC with Churchill Capital Corp. XI is pending, not closed (as of Oct 10, 2026).

Figure AI Figure 03

Figure 03 is built for volume; paid use is unclear.

Figure reports 350+ Figure 03 robots delivered by April 2026 (vendor claim); revenue and paying customers are not public.

"Delivered" includes robots used for internal R&D, data collection and home-task development, so it does not equal "in paid work".

Figure's BMW run is documented; revenue is not. Reported results, pilots and aspirations kept apart (vendor figures unless noted).

Figure discloses capacity, not price or revenue. Economics, deployment needs, labor effects and alternatives (as of Oct 2026).

Apptronik Apollo

Apollo 2 is a prototype; Apollo 3 is the product.

Apptronik pilots Apollo with Mercedes-Benz, GXO and Jabil; its first intended shipping product, Apollo 3, is expected to be unveiled in 2027.

Apollo learns tasks from people demonstrating them. A teleoperation-to-autonomy workflow, with three named pilots and no disclosed production terms.

1. Define

   Customer and Apptronik pick a task, such as moving kitted-part totes to a line side. (Customer)

2. Demonstrate

   Operators teleoperate the robot to collect demonstrations at Robot Park and customer sites. (Teleoperation)

3. Train

   Demonstrations plus simulation train AI models, including Google DeepMind Gemini Robotics. (AI models)

4. Run

   The trained policy runs on the robot; teleoperation covers gaps. (Mixed autonomy)

5. Work safely

   Force control and safety zones pause motion near people; batteries are hot-swapped. (Safety zones)

Pilots and partners (none shown beyond pilot). Robot Park workflows run at Austin, Google DeepMind, Mercedes-Benz and GXO; the number of Apollo robots built or deployed is not disclosed.

No price, no unit counts, and no shipping product. Commercial model, deployment needs and limits for Apollo (as of Oct 2026).

Valuation reports differ (about $5B to above $5.5B) and are not company-confirmed. Series A totals differ by source ($403M vs $415M initial).

Boston Dynamics Atlas

Atlas is in early production; 2026 is committed.

The electric Atlas was shown at CES on Jan 5, 2026; Boston Dynamics says 2026 units are committed to Hyundai and Google DeepMind (vendor claim).

Atlas factory use is planned for 2028, not 2026. Dated milestones and Hyundai plans; plans are not results.

A third-party caution: training activity does not establish production readiness. Hyundai's Korean union said in Jan 2026 that no robot may enter the workplace without a labor-management agreement.

Price is undisclosed; Hyundai labor terms are open. Economics, deployment needs, labor effects and open questions (as of Oct 2026).

Alternatives: Agility Digit, Apptronik Apollo, Figure 03, Tesla Optimus, XPeng IRON. Hyundai plans 25,000 Atlas units across its plants (company statement, May 2026).

Tesla Optimus

Optimus has targets and prototypes, not product.

Tesla's humanoid program has no external customers and no verified production specs as of October 2026.

Company targets are targets, not output. The reported ramp (several hundred a week in Aug 2026) rests on one paywalled source.

Optimus timelines have slipped since 2022. Company statements set against the verified outcome.

Reports conflict: Electrek (Aug 20, 2026) says only a few test robots exist, while the ramp above is one paywalled source. A "50,000 to 100,000 in 2026" target in one article does not appear in Tesla's Apr 2026 call coverage and is treated as unverified.

Price, capacity and sales dates are company targets. None of these numbers is a measured result (as of Oct 2026).

Quadrupeds are proven; humanoids are mostly pilots.

Public evidence by task, from none found to scaled or paid use (as of Oct 2026; mostly vendor-reported).

Legend: scaled or paid; pilot or early; planned; none found.

No public data center construction deployment was found for any of the seven. ANYmal data center use is one 2022 site; Spot data center use rests on 2023 reports. Spot 1,500+ and ANYmal 200+ are vendor claims; Unitree shipped about 23,000 quadrupeds in 2025. "n/d" means not disclosed.

Layout and drilling are the most commercial tasks.

The work starts from a drawing, repeats thousands of times and has a clear before-and-after measure.

Data center evidence: Dusty at Mortenson (two data centers, 700,000 sq ft); Rugged at Hensel Phelps (250,000+ sq ft); DEWALT/August pilot on 10 projects. No public Jaibot data center deployment found.

One BIM model feeds layout, drilling and as-built.

The same coordinated model sets where lines are printed, where holes are drilled and what is recorded afterward.

1. Coordinate

   Trades coordinate in Revit or AutoCAD. Vendors clean the file and resolve clashes before any robot moves. (Output: MODEL)

2. Print layout

   Dusty or Rugged prints each trade's lines, footprints and labels on the slab from the model. (Output: SLAB)

3. Drill holes

   Jaibot drills overhead anchors from the model; DEWALT/August drills holes down into the slab. (Output: ANCHORS)

4. Install

   Crews hang supports and set racks from marked or drilled points instead of measuring each one. (Output: TRADES)

5. Record

   Rugged logs completed work as as-built; Jaibot syncs progress and Field-to-BIM data to the cloud. (Output: AS-BUILT)

Dusty Robotics FieldPrinter

FieldPrinter prints full-scale layout on the slab.

1. Export

   Export the coordinated model from Revit or AutoCAD with the Dusty plugin. (BIM)

2. Prepare

   Convert it to layout files in Dusty Portal and check for conflicts before printing. (PORTAL)

3. Locate

   Set up a laser tracker on the slab so it can locate the robot. (TRACKER)

4. Print

   The robot prints each trade's layout at full scale and avoids obstacles. (FIELDPRINTER)

5. Monitor

   The operator watches progress; QR codes on the slab link to documents. (IPAD)

The workflow needs BIM files, a laser tracker with sightlines and one trained operator.

Single secondary source: a two-person crew at $75-$95 per hour loaded costs $26,000-$33,000 a month; break-even is displacing about 20% of one crew's hours.

Rugged Robotics layout robot

Rugged sells slab layout as a service.

Rugged's results are vendor-reported, not audited. Named projects come from Rugged's own pages; separate what is shown from what is still a goal.

Another vendor page claims up to 60% less layout time and 75% less rework, so figures differ by page. Rugged won the IEEE RAS Product Innovation Award at ICRA 2025.

Rugged and Dusty differ in who runs the robot. Same job, two commercial models. Status as of Oct 2026.

A third option is HP SitePrint: a partner whitepaper reports a 120,000 sq ft hyperscale site laid out in 2 days instead of 20 with two robots (vendor/partner claim).

Hilti Jaibot

Jaibot drills overhead anchor holes from BIM data.

An operator moves Jaibot; the robot drills. Workflow on the left; what has and has not been shown on the right.

Hilti also agreed in March 2023 to manufacture Canvas drywall-finishing systems on the Jaibot platform; terms undisclosed.

Hilti rents Jaibot; it does not sell it. Terms from Hilti's US Jaibot system terms (2026). Prices sit in each Order Form and are not public.

Reported price (single source, Sep 2026): $10,000+ per month, about $500 per working day. Alternative: DEWALT/August downward drilling (announced Jan 2026; vendor claims).

Raise Robotics Facade work cell

Raise builds a modular work cell for facade jobs.