Physical AI & Embodied Intelligence
The thesis that AI which perceives and acts in the physical world, through a robot, drone or autonomous machine, is its own investable layer: robot foundation models and robot learning, humanoids, mobile manipulators, autonomous vehicles and drones, intelligent robot cells, and the simulation and data platforms built specifically to train robots.
151
active companies
$30.0B
tracked funding
$24M
median raise (144 funded)
2022
median founding year
305
verified customer relationships
Membership: this thesis appears in the company’s multi-select Investment Thesis field — a company can express more than one thesis. 133 of the 151 companies here also appear in other thesis cohorts. Active companies only; exited companies are excluded.
The Thesis
This is the newest thesis in the database and, by tracked capital, the biggest. When it was introduced in November 2026 it covered about 150 active companies with roughly $30B in tracked funding: more money than the System of Record Enhancement cohort, carried by about a quarter as many companies. The cohort is also young. The median company was founded in 2022, and more than half were founded in 2022 or later. Physical AI is being funded on the premise that robot learning has finally become a software problem.
Read the aggregate carefully. Anduril alone accounts for about $7.0B of the total, and roughly a dozen companies have raised $500M or more each, so the funding figure describes a handful of infrastructure-scale bets as much as the cohort. The median raise among funded companies is about $24M, and Seed and Series A are the two largest stage buckets. Underneath the headline sits a broad early-stage layer of robot-cell, inspection and mobile-manipulation companies. The geography is American first (about two thirds of the cohort), with Germany and China the next largest bases.
How it relates to the neighboring theses: most of these companies sit on the Physical AI investment list, and a large share also carry Physics & Domain AI, because training robots is a physics and data problem before it is a hardware one. Closed-loop Automation and Lights-out Automation describe what the machines do once deployed; this thesis describes the intelligence that lets them do it. Expect the three pages to converge as the cohort matures, and treat this one as the umbrella view.
Figures reflect the ThreadMoat dataset as of November 14, 2026. The dataset updates continuously; the numbers in the sections below recompute with every data sync.
Incumbents & Evergreens
Installed base
The industrial-robot makers that own the arms, controllers and integrator channels already in the plants these startups sell into.
FANUC
Japanese maker of industrial robots and CNC controls, with one of the largest installed bases of factory robots worldwide.
ABB Robotics
Robot arms, controllers and programming software from ABB, which agreed in 2025 to sell the unit to SoftBank Group.
Yaskawa
Japanese maker of Motoman industrial robots and the motion-control drives behind them.
KUKA
German industrial-robot maker, owned by China’s Midea Group since 2017.
Universal Robots
Danish pioneer of collaborative robots (cobots), part of Teradyne.
Platform contenders
Not incumbents in the installed-base sense: the large technology companies building general-purpose robot models and platforms, as suppliers to this cohort and as competitors to it.
NVIDIA
Compute, simulation (Isaac Sim, Omniverse) and the Isaac GR00T robot foundation models much of the cohort builds on.
Google DeepMind
Gemini Robotics: general-purpose robot foundation models from the largest AI research lab.
Tesla
The Optimus humanoid program, built on its vehicle-autonomy stack and in-house manufacturing.
Amazon Robotics
Runs one of the largest mobile-robot fleets in the world inside Amazon’s own warehouses.
Boston Dynamics
Hyundai-owned maker of the Spot, Stretch and electric Atlas robots.
Autonomy & defense (examples)
A few established operators of autonomous machines outside the factory, for scale. Not a complete list.
Waymo
Alphabet’s driverless ride-hailing service, the most mature commercial autonomy stack on public roads.
Caterpillar
Autonomous haulage (Cat MineStar Command) running driverless trucks at mines.
General Atomics
Maker of the MQ-9 Reaper and other uncrewed aircraft, the established drone prime.
Trends Shaping This Thesis
Physical AI as Industrial Infrastructure
General-purpose robotics and autonomy stacks are being funded as infrastructure, not as machines. Capital requirements put this game out of reach for conventional seed-stage entrants.
Manufacturing AI Capital Concentration
Factory-floor AI is the most crowded funded category in the dataset. More capital than distinct problems usually precedes a shakeout.
The Factory Futures Funding Mirage
Heavy aggregate funding in shop-floor AI is read as market validation. Our reading: it is overcrowding, and most of that capital will not find distinct winners.
Full trend analysis with evidence and confidence grading is in the quarterly ThreadMoat report.
Top 3 Startups
Ranked by Weighted Startup Quality Score — the site-wide composite of market opportunity, team execution, technology differentiation, funding efficiency, growth, industry impact and moat. Ties break on verified customer relationships. Scores, funding and stage per company are dashboard data.
148 more ranked companies in this cohort
The full sortable list with score components lives in the ThreadMoat dashboard.
Funding Profile
Latest funding round across the cohort. 144 of 151 companies have disclosed funding above zero; the median total raise among them is $24M.
Industries Served
The industries this cohort sells into, ranked by company count (multi-select — a company serves several industries).
Customer Evidence
305 verified customer relationships across 265 organizations, joined to this cohort by record ID — never by name matching. Customer evidence exists for 68 of 151 companies; investor data for 150 of 151.
Buyer industries for these relationships, and the customer-to-startup mapping, are dashboard data.
The complete customer intelligence view — which organizations buy from which startups — is in the dashboard.
Geography
HQ country for all 151 companies in this cohort, against the full-dataset baseline.
Go deeper on this cohort
Every section above has a full interactive version in the ThreadMoat dashboard, filtered and sortable across the complete dataset.
Frequently Asked Questions
What is Physical AI & Embodied Intelligence?
AI that perceives and acts in the physical world through a machine: robot foundation models and robot learning, humanoids, mobile manipulators, autonomous vehicles and drones, intelligent robot cells, and the simulation and data platforms built specifically to train robots.
How is this different from Lights-out and Closed-loop Automation?
Those two theses describe operating models: factories that run without people, and processes that correct themselves. This thesis describes the embodied intelligence underneath, the models and robots that make autonomy possible. Many companies carry more than one of the three.
Is the funding total meaningful?
Only with context. A small number of infrastructure-scale raises, led by Anduril at about $7.0B, dominate the total, while the median raise among funded companies is around $24M. The stage mix, led by Seed and Series A, says more about the cohort than the aggregate does.
Related Investment Theses
133 of 151 companies also appear in other thesis cohorts.
Physics & Domain AI
The thesis that AI models trained on physics and deep domain data — surrogate solvers, domain foundation models, physics-informed ML — can attack the computational core of engineering software, not just its edges.
Closed-loop Automation
The thesis that the next wave of industrial autonomy closes the loop — systems that sense, decide and act without a human in each cycle, from robotic work cells to self-correcting production processes.
Agentic Orchestration
The thesis that AI agents which execute multi-step engineering and operations workflows — not just recommend actions — become the coordination layer across PLM, supply chain, manufacturing and design tools.
Lights-out Automation
The thesis that fully autonomous physical operations — factories, sites and missions that run without humans present — justify infrastructure-scale capital. Hardware-led autonomy at venture scale.
System of Record Enhancement
The thesis that the durable value in industrial AI accrues to companies that enhance the established systems of record — PLM, MES, ERP, maintenance and quality platforms — rather than trying to replace them.
AI-Enabled User Experience
The thesis that AI-native interfaces — copilots, text-to-CAD, conversational tooling, automated drawing and model interaction — can win engineering users faster than incumbents can retrofit their products.
Construction & Engineering Technology Enhancement
The thesis that construction and AEC workflows — capture, BIM coordination, field execution, building systems — reward startups that enhance the established Procore/Autodesk/Bentley stack with AI rather than replace it.
Augmented Knowledge
The thesis that AI can capture, transfer and amplify expert industrial knowledge — training, tribal-knowledge capture, guided operations — as its own product category.
Methodology
Membership. A company belongs to this cohort when “Physical AI & Embodied Intelligence” appears in its multi-select Investment Thesis field. Theses overlap: 133 of 151 members carry at least one other thesis, so cohort sums across pages exceed the database total.
Active filter. Ranked lists include active company groups only (raisers, bootstrapped, scale-ups, unicorns, corporate carveouts). Acquired, IPO’d and inactive companies appear in the exits strip; stealth and speculative outliers never appear. Unlike the site-wide software-only headline count, thesis membership keeps hardware-led autonomy companies — dropping them would erase the Lights-out and Closed-loop cohorts. Company profile links are only shown for companies in the software-only dashboard set.
Honesty rules. Percentages render only above a denominator of 30; medians only with 10+ funded companies; every breakdown states its denominator and field coverage. Customer evidence exists for 68 of 151 companies; investor data for 150 of 151.
Data vintage. Computed from the ThreadMoat dataset synced November 14, 2026. Figures recompute on every sync.