Chuhee LeeIndependent research
All publications
Abstract cover art.
Horizon
2026–2031
Scope
Global · US, China, Europe
Lens
Work, value, daily life
Basis
Evidence and forecasts
ShiftsPhysical AINovember 2026 edition · Research report

Physical AI Outlook 2026–2031

How physical AI will change work, who captures the value and daily life by 2031, with the strongest evidence and the sharpest disagreements laid side by side.

Research questionHow will physical AI, meaning AI that senses and acts in the real world, change work, the economy and everyday life by 2031, and where will it scale first?

By Chuhee Lee≈23 min readData as of 3 Nov 2026Next refresh: November 2027Change log

With research assistance from several AI models (Claude, Gemini, ChatGPT) · Views are the author’s own

Cite asChuhee Lee, Physical AI Outlook 2026–2031, November 2026 edition (data as of 3 November 2026), https://blueplanet-ai.github.io/physical-ai-outlook/, CC BY-NC 4.0.

The report in 60 seconds

Three takeaways · each with its source
  1. 5 million vs 7,000 Fact

    The robots at scale are not the ones in the headlines

    Factories now run 5 million industrial robots, and 2025 installations topped 600,000. Against that, the International Federation of Robotics counts about 7,000 full-size humanoids sold worldwide in 2025 for commercial use, and Omdia, with a wider definition, about 13,000. Forecasts for 2030 range from 500,000 to 1.2 million humanoids a year, so the range itself is the finding.

    Source: IFR [1], [2]; Omdia via TechNode [3]; Goldman [4]
  2. 74% vs 0.3% Interpretation

    What robots can do is far wider than what pays

    An Anthropic study rates robots able to perform 74% of physical tasks in the US, about 34% of working hours, but cost-competitive against human labor for only 0.3% of job tasks. Gartner expects fewer than 20 companies to run humanoids in production by 2028. Cost, safety and proof, not demos, set the pace. Anthropic wrote that study and Claude scored its tasks, so it is one input here, not the anchor.

    Source: Anthropic [5]; Gartner [6]
  3. $9.9B vs <15% Hypothesis

    Money runs ahead of use; value may settle with the few

    Robotics startups raised $9.9 billion in the third quarter of 2026, more than twice a year earlier, yet in Gasgoo’s Chinese trade data under 15% of humanoid orders were used in production work. The author’s central case, “Warehouses and factories first, homes last”, is that purpose-built machines scale first, humanoids stay narrow, and the gains go to a few chip, model and operator companies: not yet observed.

    Source: Crunchbase via RuntimeWire [7]; Gasgoo [8]

Fact: stated by the cited source. Interpretation: the author’s reading of the cited evidence. Hypothesis: expected from the evidence but not yet observed. Figures inside every takeaway are sourced.

At a glance

One-page summary

Headline findingPhysical AI will scale first where tasks are structured, the value of an hour of work is high and a failure is cheap to contain: warehouses and factories. Homes come last. Capability is wider than the economic case, so most of the next five years is about cost, safety and proof, not about what a robot can do in a demo.

Summary

The short version

Physical AI, meaning AI that senses and acts in the real world, is already large in one form and small in another. Conventional industrial and warehouse robots are a measured, growing business. The general-purpose humanoids that dominate attention number in the thousands, and forecasts for them disagree by wide margins. The author’s central case for 2031 is that machines scale first where work is structured, the value of an hour is high and a failure is cheap to contain: warehouses and factories. Farms, construction, energy and defense follow in narrower forms. Homes and care come last. Most of the gains land with a few companies in chips, models and operations, and the cost, safety and proof needed before wide use set the pace more than capability does.

The measured base is conventional robots.

The IFR counts 5 million industrial robots in operation, more than 600,000 installed in 2025 and about 250,000 professional service robots sold, of which 117,500 were logistics machines. China installed 354,000, or 59% of the world’s industrial robots. The IFR forecasts 806,000 installations in 2029.[1][2]

Humanoids are early, and forecasts split.

Counts for 2025 run from about 7,000 to 20,000 depending on definition and source. For 2026, forecasts run from 50,000 (China only) to nearly 100,000 worldwide. Gartner expects fewer than 20 companies to reach production by 2028, and Interact Analysis puts the inflection in 2032.[2][3][4][6][9]

Work changes in tasks before it changes in headcount.

Robots can technically do about a third of US working hours, but are cost-competitive for 0.3% of job tasks in Anthropic’s analysis. Studies of past robot adoption find job losses in manufacturing offset elsewhere in Germany, and local harm in US manufacturing towns.[5][10][11]

Capital arrives faster than revenue.

Robotics startups raised $9.9 billion in the third quarter of 2026 and Unitree’s shares rose 460% at the first-day close, then fell about 45% from the intraday peak within a week. Public data cannot yet show whether chips, models, makers or operators earn the margin.[7][12]

Rules arrive on dates.

The EU Machinery Regulation applies in most provisions from 20 January 2027 and treats AI safety components that change their own behavior as high-risk. A revised ISO standard for industrial robot safety was cited in the EU Official Journal in September 2026. As of September 2025, UN talks on autonomous weapons had not reached negotiations; the 2026 outcome is unconfirmed.[13][14][15]

The forecast spread: humanoid shipments in 2030

Thousand units a year · bank and analyst forecasts

Smart Analytics Globalworld
500,000
Morgan StanleyChina only
446,000
Goldman Sachsworld
890,000
Bank of Americaworld
1.2 million
Interact Analysisworld, 2035 only
700,000+ in 2035

Different geographies and publication dates; Morgan Stanley covers China only, and Interact Analysis forecasts 2035, not 2030. Goldman raised its 2030 figure from 256,000 this year.[4][16][9][17] The author reads the width of the range as the main fact.

Chapter 1 of 9
Baseline

Where we start in 2026

Physical AI is measured today by conventional robots; the machines in the headlines are a rounding error next to them, and every count is a proxy for what this report is about.

Definition. Physical AI here means AI that senses the world and acts on it under safety constraints. No official statistic counts that. The International Federation of Robotics (IFR) counts robots by type, whether or not a learned model controls them, so every figure below is a proxy for adoption, not a count of physical AI itself.

Factories. The IFR counts 5 million industrial robots in operation, up 9%, and more than 600,000 installed in 2025, up 11%. China installed 354,000 (59% of the world), the United States almost 38,500 (up 12%), Japan 36,219 (down 19%), South Korea 30,000 and Germany under 25,000 (down 8%). Chinese suppliers made 195,000 of China’s robots, a 55% domestic share.[1] The IFR forecasts 655,000 installations in 2026 and 806,000 in 2029, and its release does not mention humanoids.[1]

Services and homes. Professional service robots sold almost 250,000 units in 2025, up 24%, with logistics robots the largest group at 117,500 (47%). Consumer robots, mostly floor cleaners and lawn mowers, sold almost 34.2 million units, up 37%.[2] Amazon said in June 2025 that it had deployed its millionth robot.[18]

Humanoids. The IFR counts about 7,000 full-size humanoids (above 140 cm) sold for commercial use in 2025.[2] Omdia estimated about 13,000 shipped in 2025 (published January 2026), led by AgiBot (5,168), Unitree (4,200) and UBTech (1,000).[3] Bank of America puts 2025 shipments at 20,000, again on its own definition.[17] The two figures use different definitions and should not be averaged or added to the industrial count. Unitree’s prospectus gives 2025 revenue of 1.699 billion yuan and “over 5,500” humanoids sold.[19]

Driving as a proof point. Driving is the one place where learned systems already act on public roads at consumer scale, and it is covered in the site’s mobility reports rather than here. Tesla reports 1.74 million active FSD subscriptions in the third quarter of 2026, a company figure relayed by a secondary source.[20] It shows what a structured, high-value, rule-bound task looks like when it scales.

Industrial robots installed in 2025, by country

Thousand units · IFR World Robotics 2026

China59% of the world
354,000
United States+12%
almost 38,500
Japan−19%
36,219
South Korea−1%
30,000
Germany−8%
under 25,000

Germany is drawn at 25,000; the IFR reports “fewer than 25,000”.[1]

Humanoids in 2025: two definitions, one order of magnitude

Units · IFR (sold, full-size) and Omdia (shipped)

Omdia, all humanoidsanalyst estimate
about 13,000
IFR, full-size humanoidsabove 140 cm, commercial use
about 7,000
AgiBotOmdia
5,168
UnitreeOmdia
4,200
UBTechOmdia
1,000

Company bars are Omdia’s split of its own total. Counts of shipments include units bought for research, demonstration and entertainment.[2][3]

Author’s read · confidence: highPhysical AI starts 2026 as a story of two scales: millions of conventional machines doing fixed work, and thousands of learning-based general-purpose machines being tested. The five-year question is how much of the first becomes the second, and whether it pays.
Chapter 2 of 9
Layer 1 · models and silicon

The models and chips under the robot

Robot foundation models and edge computers are improving fast, but the independent evidence says the gap between a controlled demo and the real world is still wide.

What is claimed. NVIDIA calls its Jetson Thor edge computer “the ultimate supercomputer to drive the age of physical AI and general robotics”, a product claim with no independent deployment data behind it.[21] Model developers argue that large vision-language-action models trained across many robot types will let one model control many machines.

What independent work finds. The Stanford AI Index 2026 finds that robots still fail at most household tasks even as they excel in controlled environments.[22] Robot training data must come from physical robots or simulation, which is why it is scarcer than the text and images behind chatbots.

What is not known. No source opened for this report gives audited uptime, intervention rates or maintenance cost for learned-model robots in commercial service.

What cuts against this. Models can improve faster than hardware or data collection suggests, and edge computers are getting cheaper. A model that transfers between machines would lower the cost of every new task. Nothing in the evidence collected here shows that transfer working across body types and sites at commercial reliability.
Author’s forecast · confidence: mediumBy 2031 learned models will do the perception and the flexible handling inside structured cells, such as picking varied items from bins, while safety-critical control stays deterministic. A single model that runs any machine in any setting is not the base case.
Chapter 3 of 9
Sector 1 · factories and warehouses

Factories and warehouses

The purpose-built machines scale; the humanoids so far mostly demonstrate. The next five years decide whether any of them earn a place on the floor.

The measured part. Warehouse automation is already commercial. Amazon said in 2025 it had deployed its millionth robot; Symbotic reported $721 million of quarterly revenue and 77 systems in deployment.[18][23] Logistics is the largest group of professional service robots at 117,500 units a year.[2] Product selection still fails: Amazon stopped its Blue Jay warehouse robot after under six months while keeping its other robots.[24]

Humanoids on the floor. Boston Dynamics said in January that production Atlas robots begin manufacturing immediately and that all 2026 units are committed to Hyundai’s robotics application center and Google DeepMind, with more customers in early 2027. Hyundai announced a $26 billion US investment that includes a robot factory for 30,000 robots a year.[25] Tesla says Optimus production has started on its Fremont lines and gave no unit count; Elon Musk has said Tesla could make 20,000 a week by the end of 2027, a target, not output.[26] All of these are company statements.

Who buys. In Gasgoo’s trade data on Chinese humanoids, 35% of shipments went to entertainment and performance, 20% to data collection and research, and under 15% of orders were used in production work that creates direct industrial value; Gasgoo also saw no large repeat orders or proven return-on-investment model. Prices range from under 10,000 yuan to more than 400,000 yuan.[8] Gartner expects “fewer than 20 companies” to go live in production with humanoids in supply chain and manufacturing through 2028, and says they cost multiple times more than task-specific robots with lower throughput.[6]

Forecasts. Forecasts of humanoid units are unsettled. For 2026 they run from 50,000 in China (Morgan Stanley) to nearly 100,000 worldwide (Gasgoo), a 1.7-times spread among the global estimates. For 2030 they run from 500,000 (Smart Analytics Global) to 890,000 (Goldman Sachs), and Bank of America reaches 1.2 million.[16][8][4][17] Goldman raised its 2030 figure from 256,000 this year. Interact Analysis forecasts over 700,000 units and $15 billion of revenue only in 2035, with real-world applications about a tenth of 2025 production.[4][9]

Humanoid shipments in 2026: the forecasts

Thousand units · bank and analyst forecasts

Morgan StanleyChina only
50,000
Smart Analytics Globalworld
~60,000
Goldman Sachsworld
75,000
Bank of Americaworld
90,000
Gasgoo researchworld
~100,000

Morgan Stanley’s figure covers China only and is not comparable with the global ones.[16][8][4][17]

What buyers use humanoids for

Share of shipments and orders · Gasgoo, August 2026

Entertainment and performance
35%
Data collection and research
20%
Production work with direct industrial value
under 15%

The outlet’s own trade data; method not shown. The bars do not sum to 100%.[8]

What cuts against this. Brownfield sites are built for people, and a human-shaped machine needs no re-tooling; that is the case for humanoids. Purpose-built robots give up that flexibility but are cheaper and faster, which is why Gartner and Interact Analysis lean toward them. Humanoid list prices in China already start under 10,000 yuan,[8] and the first named customers for Atlas are Hyundai and Google DeepMind.[25]
Author’s forecast · confidence: mediumBy 2031 warehouses and factories are where physical AI is most visible: more mobile and fixed robots with learned perception, a few thousand humanoids doing tote handling, parts moving and inspection in big-company sites, and no general-purpose factory worker. Industrial installations keep rising, near the IFR’s 806,000 a year in 2029.
Chapter 4 of 9
Sector 2 · homes, care and healthcare

Homes, care and healthcare

The home is the hardest setting for a machine and the last to be won. Healthcare already uses robots, but the surgeon stays in the loop.

Homes. Consumer robot sales of almost 34.2 million units in 2025 are floor cleaners, lawn mowers and similar narrow machines.[2] They show demand for a single useful job at a low price, not for a general household helper. The opened sources contain no probability-sample survey of willingness to pay for a home humanoid, and no audited home deployment, so those are gaps. Interact Analysis lists safety and environmental complexity as the barrier for household use.[9]

Surgery. Intuitive Surgical’s da Vinci systems supported about 3.15 million procedures in 2025 (preliminary, unaudited), and the company expects 13% to 15% growth in 2026.[27] The IFR counted 8,400 surgical robots sold.[2] The FDA states that these devices cannot operate without direct human control.[28] Growth in surgical robots is therefore growth in surgeon tools, not in autonomous care.

Care. The research found no commercial deployment data on elder-care humanoids and no sourced survey of hospital or home-care buyers; claims that hospitals will or will not adopt them are unverified.

What cuts against this. Labor shortages, which the IFR names as a driver of factory automation, are also felt in care,[1] and a robot that lifts or fetches could be worth far more there than in a factory. A cheap, safe single-purpose device, such as a patient-transfer aid, could be adopted well before any general helper. The evidence collected here does not show either.
Author’s forecast · confidence: medium for surgery, low for homesBy 2031 the home still holds single-purpose robots, now with better perception, and healthcare holds more surgeon-guided systems and more hospital logistics robots. A general-purpose home or care humanoid sold in volume to ordinary households is not the base case.
Chapter 5 of 9
Sector 3 · farms, construction, energy and defense

Farms, construction, energy and defense

Outdoor and high-stakes settings are big labor pools but hard for machines; drones and defense move first because buyers there pay for capability, not payback.

Thin evidence. This is the least documented chapter. The research found no current, comparable deployment dataset for agriculture, construction, mining or utilities. The numbers that exist are about money and rules, not machines in the field.

Money. Crunchbase’s article on robotics funding to mid-June 2026 names a $1.75 billion Series D for Saronic, a defense maritime company, valued at $9.25 billion; it does not say whether the $18.8 billion total includes defense.[29] One secondary account says the third-quarter total of $9.9 billion excludes defense companies, while another does not say.[7][30] The counts may therefore not be comparable. Drone unit volumes are not sourced anywhere in this report.

Rules. The International Committee of the Red Cross separates weapons that merely fly themselves from those that select and apply force to targets, and says legal compliance may require certain autonomous weapons to be ruled out.[31] At the UN’s expert group on autonomous weapons in September 2025, 42 states said they were ready to move toward negotiations on the chair’s draft text. Negotiations had not begun. Whether the 2026 session moved further is not confirmed in the sources opened.[15]

What cuts against this. Mud, dust, weather and lost satellite signals make the field harder than a warehouse, and construction sites mix machines with tradespeople under strict liability. Defense procurement follows conflicts and budgets, not commercial payback, so what it shows does not transfer to industry.
Author’s forecast · confidence: lowBy 2031 drones, autonomous field machinery with a human supervisor, and inspection robots are the growth in this group; machines that manipulate tools on a building site or farm remain pilots. Defense is the fastest-growing buyer, and it is also where public rules are least settled.
Chapter 6 of 9
Impact 1 · work and labor

Work and labor

Robots are already able to do a third of US working hours on paper. The economics, and the history of earlier robots, point to slower and more local change than the headlines imply.

The new estimate. Anthropic’s Economics team rated about 7,600 physical tasks from the US occupational database. It finds robots can already perform 74% of physical tasks, about 34% of working hours, mostly in limited settings. It estimates that robots are cost-competitive against human labor for just 0.3% of job tasks. In its worked example the cost model assumes a ten-year service life and an 8% cost of capital, and it assumes robot prices falling about 3% a year, a pace at which 10% cost-competitiveness would take about 40 years. The authors call the result suggestive and note that AI could let robots learn tasks they cannot do today.[5] Anthropic wrote the study and Claude scored the tasks. This report is also produced with Claude, so the figure is one input among several, not the anchor.

What earlier robots did. A German study of 1994–2014 found each industrial robot destroyed about two manufacturing jobs, with the loss fully offset by additional jobs in services. Incumbent workers in exposed plants mostly stayed with their employer, at lower wages; the manufacturing decline came from fewer hires of young entrants.[10] In the United States, one more robot per thousand workers lowered the local employment rate by about 0.2 percentage points and wages by about 0.42% across commuting zones between 1990 and 2007.[11] A later German study found that works councils reduce separations among incumbent workers and do not hinder robot adoption.[32]

What forecasters add. McKinsey’s US scenario to 2035 sees automation reducing labor demand by the equivalent of about 36 million jobs while other demand creates over 40 million, but it covers all AI and is not about robots alone.[33] The IFR points to labor shortages, demographic change and the relocation of manufacturing as drivers of factory automation.[1] No source found models jobs created or lost by physical AI to 2031, and this report offers no net-jobs number.

Able, not yet worth it

Share of US physical tasks, working hours and job tasks · Anthropic Economics, September 2026

Physical tasks robots can already doshare of tasks
74%
Share of US working hours those tasks make up
34%
Job tasks where a robot is cost-competitiveshare of tasks
0.3%

The first two bars measure technical capability, the third cost-competitiveness; the bars have different bases.[5]

What cuts against this. The cost result depends on a flat 3% yearly price decline for every task, which the authors say may not match how robot makers choose what to improve. Humanoid list prices in China already start under 10,000 yuan.[8] Past studies cover fixed industrial arms, not learning machines that could change tasks across occupations.
Author’s forecast · confidence: mediumBy 2031 the labor effect is concentrated: fewer repetitive handling and assembly tasks in the sites that adopt, more supervisor, technician and exception-handling work around them, and local effects in manufacturing and logistics towns. Economy-wide employment does not move visibly because of robots.
Chapter 7 of 9
Impact 2 · economy and value

Who captures the value

Money is flowing in at record speed, but public data cannot yet show which layer earns the margin: chips, models, makers or operators.

Capital. Robotics startups raised $18.8 billion in 2026 through mid-June, more than the $15 billion of all of 2025 and the $14.1 billion of 2021, and $9.9 billion in the third quarter alone, more than twice the $4.3 billion of a year earlier. Year to date, Chinese robotics startups raised $13.2 billion and US ones $12.1 billion.[29][7] Funding is capital raised, not sales.

Public markets. Unitree listed on Shanghai’s STAR Market on 19 August 2026, rose 460% at the first-day close, with a value touching about $66 billion intraday, then fell about 45% from that peak by 25 August.[12] Reuters reported on 28 October that the shares had fallen further, almost 55% from their debut, before a partial rebound.[34] For the first nine months of 2026 its revenue rose 55.7% to 1.817 billion yuan, but profit excluding one-off items fell 18.8% as research spending rose 191%. Reuters reported third-quarter profit down 17.4%.[35][34] Goldman Sachs projects a $138 billion humanoid market in 2035.[4]

The open question. The working hypothesis is that value concentrates in chips, models and a few operators. The opened sources do not support or reject it: no public series compares gross margins along the stack, and the reviewers who tested it disagree. Chip and model vendors cite scarce inputs; system integrators and service firms argue that safety certification, spare parts and uptime guarantees are where customers stay locked in.

Three setups. China has manufacturing scale: Chinese suppliers hold 55% of their home market and the country installed 59% of the world’s industrial robots.[1] The United States has cloud, model and operating capital, with Amazon, Tesla and the major chip designers. Europe has automation incumbents and the first hard rules. Appendix A lists named players.

Robotics startup funding

Billion dollars · Crunchbase

Full year 2021
$14.1B
Full year 2025
$15.0B
2026 through mid-June
$18.8B
Q3 2026 aloneseparate count, defense treatment unclear
$9.9B

Different periods and category definitions. Whether defense is included is not stated for the June count, and one account says the Q3 count excludes it.[29][7]

What cuts against this. A crowded, subsidized field can compress margins quickly, and a listing priced on hope can reverse in days, as Unitree’s did. A winner-take-most outcome in robot models, as in cloud, could still give upstream suppliers the largest share.
Author’s read · confidence: lowBy 2031 a handful of firms in chips, robot models and large-scale operation will hold most of the visible profit, and many of today’s hardware startups will have merged, pivoted or closed. The evidence for that is the pattern in earlier hardware-and-software markets, not yet a measurement in robotics.
Chapter 8 of 9
Counterforces

What holds it back

Seven forces slow the shifts: a real-world gap in models, cost, proof, safety rules, weapons law, capital cycles and measurement.

ForceEvidenceEffect on the shiftsKind
Real-world gapStanford’s AI Index: robots still fail at most household tasks even as they excel in controlled environments.[22]Keeps learned models inside structured cells.Survey
CostCost-competitive for 0.3% of job tasks; Gartner: humanoids cost multiple times more than task-specific robots.[5][6]Slows adoption to tasks with high labor cost per hour.Forecast
ProofUnder 15% of humanoid orders used in production work; real-world applications about a tenth of 2025 production.[8][9]Buyers wait for repeat orders and audited return on investment.Observed
Safety rulesEU Machinery Regulation applies from 20 January 2027 (most provisions), with AI safety components that change their own behavior on the high-risk list; ISO 10218:2025 replaces the blanket safety-performance requirement with function-specific levels, and European references were published in September 2026.[13][14]Certification becomes a gate in Europe, and sets design norms elsewhere.Observed
Weapons lawUN experts had a draft text and 42 states ready to negotiate in September 2025; ICRC says autonomous targeting raises legal concerns.[15][31]Slows autonomous lethal systems; leaves drones for sensing and logistics freer.Observed
Capital cyclesUnitree fell about 45% from its intraday peak within a week of listing, and Reuters later reported a further fall and a partial rebound.[12][34]A funding reversal would hit startups before revenue catches up.Observed
MeasurementHumanoid counts differ by definition (7,000 or 13,000), and it is unclear whether funding counts include defense.[2][3][29]Makes progress hard to verify; favors claims.Observed
Author’s read · confidence: mediumNone of these forces stops physical AI. Together they make it arrive in narrower, safer forms and later than company and bank forecasts suggest, and they make Europe the first place where the safety case must be written down.
Chapter 9 of 9
Judgment, not forecast

Scenarios to 2031

The author’s central case is that structured work scales first and homes come last. Two alternatives move the pace; a stall would slow all of it.

How good are such forecasts? Humanoid forecasts have moved fast. Goldman Sachs raised its 2030 figure from 256,000 to 890,000 units and its 2035 figure from 1.38 million to 6.5 million this year; Morgan Stanley raised its China figure for 2026 from 14,000 in January to 28,000 and then to 50,000 by June.[4][16] Forecasts that move this much, with no order book behind them, deserve little weight, which is why this report shows ranges. The IFR’s industrial forecasts rest on a long record of census data and deserve more.

ScenarioWhat 2031 looks likeEvidence forEvidence againstConfidence
Warehouses and factories first, homes last
central case
Purpose-built and learned-perception robots keep growing in logistics and factories; humanoids work in a few thousand big-company sites; homes hold single-purpose machines.5 million robots in operation, 806,000 installs forecast for 2029, 117,500 logistics robots, Symbotic and Amazon fleets.Amazon stopped Blue Jay; under 15% of humanoid orders for production.Medium
General-purpose breakthroughOne robot model works across many machines and sites; humanoids reach factories, hospitals and some homes at volume.Goldman and Bank of America curves; falling Chinese prices; Atlas and DeepMind integration.Stanford’s finding on household tasks; Gartner’s fewer-than-20; Interact Analysis puts the inflection in 2032.Low
Narrow machines winWheeled and fixed robots with better AI take almost all the work; humanoids stay a niche even in factories.Gartner’s cost and throughput finding; logistics growth; surgical robots.Brownfield sites built for people; model transfer improving.Medium
StallFunding falls before revenue rises; incidents and rules slow deployment; the story resets.Unitree’s drop; order mix; profit pressure at the listed makers.Record installs; labor shortages.Low to medium

Confidence describes how well the evidence supports each story, not a probability. The central case and “Narrow machines win” overlap: they differ on whether humanoids reach production at all.

What would change the picture

Each September

Industrial installs and the service-robot count

The IFR’s World Robotics. Installs near 806,000 in 2029 support the central case; flat installs support a stall.[1][2]

Half-yearly

Humanoid shipments and the share used for production

Omdia, Interact Analysis and trade trackers. A production share well above 15% supports a breakthrough; a widening gap between shipments and use supports a stall.[3][8]

Quarterly

Optimus units and the Atlas fleet

Tesla’s updates and Hyundai’s reports. Unit counts and customers beyond Hyundai and Google DeepMind in 2027 would move the humanoid case.[26][25]

20 January 2027

EU Machinery Regulation

Guidance and the first conformity assessments for AI safety components. Smooth certification supports the central case; delays support a stall in Europe.[13]

Quarterly

Funding against revenue

Crunchbase counts against filings of listed makers such as Unitree. Funding falling before revenue rises would support a stall.[7][35]

Quarterly

Surgical procedures and installed base

Intuitive’s filings. Growth continues to show healthcare robots as surgeon tools, not autonomous labor.[27][28]

Appendices

Players and rules

Appendix A. Who is building what

PlayerSetupWhat the sources showKind
UnitreeChina2025 revenue 1.699 billion yuan; over 5,500 humanoids sold; listed 19 August 2026; Q3 revenue 665 million yuan, up 69.9%.[19][35]Company
AgiBot, UBTechChinaOmdia counts 5,168 AgiBot and 1,000 UBTech humanoids shipped in 2025.[3]Observed
Tesla (Optimus)USProduction started on Fremont lines; no unit count.[26]Company
Boston Dynamics, HyundaiUS, Korea2026 Atlas units committed to Hyundai and Google DeepMind; $26 billion US investment including a 30,000-robot-a-year factory.[25]Company
AmazonUSOne million robots deployed by June 2025; Blue Jay stopped in 2026.[18][24]Company
SymboticUS$721 million quarterly revenue and 77 systems in deployment (fiscal Q3 2026).[23]Company
Intuitive SurgicalUSAbout 3.15 million da Vinci procedures in 2025; expects 13% to 15% growth in 2026.[27]Company
NVIDIAUSJetson Thor edge computer marketed for physical AI; no independent deployment data.[21]Company
SaronicUS$1.75 billion Series D at $9.25 billion valuation (March 2026), a defense maritime company.[29]Observed

Not found in the opened sources: current status of Figure, Apptronik, Agility and 1X; the close of any Physical Intelligence round. They are left out rather than guessed.

Appendix B. Rules and standards

RuleRegionWhat it doesStatus
Machinery Regulation (EU) 2023/1230EUUpdates the Machinery Directive 2006/42/EC; puts AI safety components with self-evolving behavior on the high-risk list.[13]Most provisions apply 20 January 2027
ISO 10218-1 and -2:2025Global, EU referencesFolds ISO/TS 15066 into the robot-system part; function-specific performance levels; a separate normal-stop function. The explainer cited does not discuss humanoids or mobile manipulators.[14]Published Feb 2025; EU references Sep 2026
FDA oversight of robotic surgeryUSClears devices that operate under direct human control.[28]In force
UN expert group on autonomous weaponsGlobalChair’s draft text; 42 states ready to negotiate (Sep 2025).[15]No treaty; 2026 outcome unconfirmed
US and China equivalentsUS, ChinaNo single binding regime comparable to the EU’s was found in the research; product-by-product rules apply.Gap

Method, limits and flags

Read before citing

Question. How will physical AI, meaning AI that senses and acts in the real world, change work, the economy and everyday life by 2031, and where will it scale first?

Scope. Global, with the United States, China and Europe as three setups. Three sectors (factories and warehouses; homes, care and healthcare; farms, construction, energy and defense including drones) and three impacts (work and labor, economy and value, daily life). Driving appears only as a proof point; the site’s mobility reports cover cars and cabins.

Sources. Research papers and institutes first, then industry bodies, analysts and banks, then company disclosures and major news and trade publications. Each reference carries its type and date. Observed, survey, forecast and company figures are kept apart and labeled.

How the research was done. Claude, Gemini (papers, standards, government) and ChatGPT (analysts, filings, news) each answered the same written brief. Claude compared the three, opened the sources behind the figures used, and merged them; Gemini’s unsupported figures were dropped. Framing, the nine chapters, the scenarios and conclusions are the author’s.

What this report is not. The “author’s forecast” boxes and the four scenarios are judgments based on the evidence cited. Confidence labels describe how well the evidence supports a statement, not a probability. Signposts give a direction and a source, not a numeric threshold.

Flags and open questions

  1. Forecasts, not measurements. Banks (Goldman, Morgan Stanley, Bank of America) cover robot and chip stocks, Gartner and Interact Analysis sell analysis, and the IFR represents robot makers. Company figures (Unitree, Tesla, Boston Dynamics, Amazon, NVIDIA) are claims. Each is labeled.
  2. Research round with Gemini and ChatGPT. Both models answered the same written brief. Claude compared the three answers and opened the sources behind the figures used here. ChatGPT’s answer was largely confirmed; Gemini’s was mixed, with some precise figures that no source supported, which are left out.
  3. Corrections from that round. Gemini’s industrial-robot figures (603,307 installs; China 54%) repeated numbers from the earlier edition; the IFR page says “more than 600,000” and China 59%, so those are used. The IFR counts about 7,000 full-size humanoids and Omdia about 13,000 because the definitions differ. Gartner’s prediction is “fewer than 20 companies going live in production”, with no unit volume attached. Unitree listed on 19 August, not in July as the October edition said.
  4. The Anthropic study. The robot task and cost study was written by Anthropic’s Economics team and Claude scored the tasks. This report is also produced with Claude. The study is used as one input, with its own caveats, and no conclusion rests on it alone.
  5. Definitions differ. Industrial robots (IFR), service robots (IFR) and humanoids (IFR, Omdia) are different counts and are never added. Crunchbase’s “robotics” counts differ on defense; the 185% growth rate it publicly states is not used because the article ties it to different periods.
  6. Unitree profit. TechNode reports nine-month profit excluding one-off items down 18.8% and reported profit up 270.5% because of a one-off share-based payment a year earlier; Reuters reports third-quarter profit down 17.4%. The report quotes both and does not merge them.
  7. Independent fact-check. A reviewer who did not write the report checked its sourced claims against the cited pages on 3 November. Findings were applied: one claim about German workers was corrected, several quotations were removed or reworded, and stale or overreaching wording on Unitree, Interact Analysis, Tesla, the UN talks and defense funding was fixed. Bank of America, Symbotic, Intuitive, the IFR surgical count, the FDA statement, the IAB, ZEW and Acemoglu and Restrepo papers, the ICRC statement and Reuters’ Unitree profit figure were confirmed. McKinsey’s scenario was confirmed only through secondary reports. The Stanford AI Index is paraphrased from its confirmed highlight; the quotation first used could not be found.
  8. Secondary reports. Where the primary page could not be opened, a named secondary report is cited: Goldman and Morgan Stanley forecasts via Humanoid Guide and AI Weekly; Morgan Stanley’s 2030 China figure of 446,000 is attributed by AI Weekly’s headline and was not matched to the original CNBC report; Tesla’s update via TechCrunch and TeslaNorth; Unitree’s prospectus via Gasgoo and its listing via The Next Web; Gartner’s prediction was read on Gartner’s own page.
  9. Not found. No source found for: audited uptime and maintenance cost of humanoids in commercial service; survey data on willingness to pay for home or care robots; current deployment data for farms, construction, mining or utilities; drone unit volumes; a US or Chinese rule comparable to the EU Machinery Regulation; a 2026 outcome of the UN talks on autonomous weapons; margins along the robotics value chain; a forecast of net jobs from physical AI to 2031.
  10. Left out. A headquarters map and a full player table are not included; Appendix A lists only players with a sourced figure. Cabin and passenger-car material is covered by the site’s mobility reports.

Change log

  • November 2026 edition
  • Changed
    Replaces the October 2026 edition, “Physical AI and the Rider’s Cabin”, with a five-year forecast of physical AI across sectors. The cabin material moves to the site’s mobility reports. The address changed from /physical-ai-cabin/ to /physical-ai-outlook/. Corrections to the earlier figures are listed in the flags above.

About this report

Personal research by Chuhee Lee. Research, source verification and drafting were carried out with several AI models, including Claude (Anthropic), Gemini (Google) and ChatGPT (OpenAI). Every factual claim is sourced to the references listed below; topic selection, framing and conclusions are the author’s. These are personal views, not those of any employer.

References

Continue down the map

ForcesShiftsSignalsOpportunities
Follow new editions

Each report is refreshed on a schedule. Know when it changes.

Next refresh dates are listed on the home page.