OpenMind Physical AI Readiness Index 2026

A geospatial analysis of where on earth advanced embodied AI can be built, powered, staffed, supplied — and permitted to work.

Two composites over 58 countries:

• PAIRI — OpenMind Physical AI Readiness Index. Capability to build and operate advanced embodied AI at scale. Six weighted pillars, 0–100, higher is better.
• OLI — Operating Latitude Index. How much room law, labour and public opinion leave an operator to deploy and run autonomous machines. 0–100, higher is more permissive. (Previously the Deployment Eagerness Index; identical scores.)
• EFFECTIVE — PAIRI × (0.45 + 0.55 × OLI/100). Latitude-weighted near-term deployment potential.

For both indices, higher is better.

Headline result

#	Country	PAIRI	OLI	Effective
1	China	82.9	72	70.1
2	Japan	69.6	61	54.6
3	United States	69.0	48	49.3
4	South Korea	66.9	41	45.2
5	Germany	66.1	28	39.9
6	Taiwan	55.4	58	42.6
7	France	51.8	24	30.1
8	Switzerland	51.3	60	40.0
9	Sweden	50.6	27	30.3
10	Italy	50.5	17	27.4

Contents

OpenMind_Physical_AI_Readiness_Index_2026.pdf   25-page preprint (two-column, REVTeX-style)
src/out/
  pairi_2026.csv           full dataset, 58 countries
  pairi_clusters_2026.csv  22 metro clusters, same six pillars
  pairi_uncertainty.csv    Monte Carlo intervals per country
maps/                                  PNG + SVG figures
  map_readiness      PAIRI choropleth
  map_latitude      OLI choropleth
  map_effective      latitude-weighted potential
  map_clusters       22 metro clusters, scored
  map_pillars        six-pillar small multiples
  chart_quadrant     capability vs latitude scatter
  chart_top25        pillar decomposition, top 25
  chart_uncertainty  published score and 90% interval, top 25
src/
  report.md          <- THE PROSE. Every word of the paper, front matter
                        (title, byline, abstract, keywords, colophon) in the
                        "::: front" block at the top, body below it.
  references.md      <- THE SOURCES. One keyed record each; format at the top
  build.py           one command: rebuilds data, figures, and both PDFs
  browser.py         locates headless Chrome for the PDF step
  geodata.py         fetches and caches the country boundaries for the maps
  front.py           parses the front-matter block (format documented at the top)
  refs.py            parses references.md and numbers the citations
  md_to_html.py      report.md -> report.html (syntax documented at the top)
  preprint.py        applies the two-column preprint layout and citations
  build_preprint.py  renders the preprint PDF
  build_index.py     dataset construction + index computation
  bootstrap.py       Monte Carlo uncertainty over the analyst-assigned inputs
  make_maps.py       all figures
METHODOLOGY.md       pillars, weights, normalisation, data-quality policy

Pillars and weights

Pillar	Weight	Core question
Energy & grid	15%	Is power abundant, cheap, reliable and connectable?
Compute & semiconductors	15%	Training, edge inference, and the silicon inside a robot
Robotics & actuation	25%	Integration capacity, OEMs, and the actuator supply chain
Materials & supply chain	15%	Magnets, rare earths, cells, bearings, machine tools
Human capital & software	20%	Robot-learning researchers and integration engineers
Demand pull	10%	Shrinking workforce × high wages × industrial base

Operating latitude is scored on four 25-point components: law, labour, public opinion, and demonstrated autonomy permissiveness. 25 means maximally permissive on that dimension.

Editing the report

src/report.md is the only file that holds the report's words, and nothing in the build overwrites it. Change a sentence there and rebuild:

uv run src/build.py report     # prose -> both PDFs, skipping data and figures
uv run src/build.py            # everything, including the index and the maps

The build writes everything to src/out/ and then copies both PDFs to the repo root and all eight figures (PNG and SVG) to maps/, so the top-level artifacts are always the latest build rather than hand-placed copies. The PDFs are rendered by headless Chrome. Do not edit report.html, report_preprint.html or the PDFs: every build regenerates them.

Two things to know when rewriting:

• The ::: front block at the top of report.md holds the title, byline, abstract, keywords and colophon. Format is documented at the top of src/front.py; body syntax at the top of src/md_to_html.py.

• Citations live in the prose as keys, so you can rewrite any sentence freely and its sources travel with it:

  ... more than all other countries combined.[@ifr-world-robotics-2025]
  ... on both counts.[@iea-rare-earth-elements-2026; @adamas-2026]
  

  Numbers are assigned at build time in order of first appearance, so you never write one and never renumber. A key with no record in references.md stops the build and names the key. Renaming a key is a find-and-replace across the two files.

Watch the numbers the build reports. citations: 89 sources cited at 101 points should only change when you mean it to. links: 173 external, 124 internal means the reference URLs are still clickable in the PDF and the [n] markers still jump to the bibliography; the build warns if either falls to zero.

Adding a source

Add a record to references.md and cite its key from report.md:

[iea-rare-earth-elements-2026]
  International Energy Agency, *Rare Earth Elements*, IEA, Paris, April 2026.
  https://www.iea.org/reports/rare-earth-elements
  note: subscription        # optional; shown in italics after the address

A record that is never cited still appears, in a Background sources section at the end of the bibliography.

Reproducing

From a clean clone, one command does everything — about 80 seconds, most of it the Monte Carlo:

uv run src/build.py

It creates src/out/, fetches the Natural Earth 50m admin-0 boundaries into src/data/ on first run (pinned to tag v5.1.2, so the maps do not change when upstream edits a border), runs the index and the bootstrap, draws every figure, and renders both PDFs. Both directories are derived and are not in version control.

Metro clusters

Twenty-two clusters scored on the same six pillars. Top ten: Shenzhen–Dongguan–Guangzhou 87 · Shanghai–Suzhou–Hangzhou 86 · Tokyo–Kanagawa 80 · Seoul–Gyeonggi+Ulsan 78 · Beijing 74 · Hefei–Wuhu 72 · Nagoya–Toyota 71 · Hsinchu–Taichung–Taipei 70 · Bay Area 70 · Stuttgart–Karlsruhe 70.

pairi_clusters_2026.csv also carries q36, a separate qualitative rank on capacity to build at volume within 36 months. It disagrees with the index in two informative places: the Bay Area (3rd on q36, 9th on PAIRI) and Beijing (10th on q36, 5th on PAIRI).

Caveats

Scores are analytical constructs, not measurements. Roughly 55% of the weighted index rests on cited figures; the remainder are analyst scores anchored to documented evidence, because the underlying data does not exist publicly — no country publishes a count of robot system integrators, IFR publishes robot density for only ~14 countries free, and sensor market shares are paywalled. Differences of under about 9 points between countries carry no ordering information — measured, not asserted, by a Monte Carlo (bootstrap.py) that models correlated single-rater bias and resamples the pillar weights on every draw. Scores were assigned by a single analyst with no second rater, and OpenMind has a commercial interest in this subject; both are stated in the report. See §12 of the report and METHODOLOGY.md for the full gap list.

Published by OpenMind, openmind.com. Issued August 2026. Data current as of August 2026. Questions and corrections: research@openmind.com.