# Methodology

## What We Select

B2AGI Physical AI Ranking 2027 examines how intelligence reaches physical work. The edition contains 20 countries and economies, 20 cities and industrial clusters, 200 companies, 100 people and 50 institutions.

Physical AI connects perception, learning, planning or adaptive control to action in a physical environment. Components, development tools, operators, capital providers and institutions enter the selection through a documented contribution to that chain. Each rationale identifies the contribution being recognized.

## Five Analytical Lenses

| Lens | Assessment question | Evidence we examine |
| --- | --- | --- |
| Intelligence | What does the system perceive, learn, plan or control? | Technical papers, model and software releases, control methods and their demonstrated tasks |
| Physical | What work reaches a robot, component or customer site? | Physical tests, customer trials, integration records and operating reports |
| Capital | How is development or deployment funded? | Financing, customer contracts, capital-market access and financial statements |
| Trust | Who defines and checks the operating conditions? | Supervision, test protocols, certificates, standards and regulatory decisions |
| Beyond | How does the ecosystem invest in its next generation of capabilities? | R&D, education and fixed-capital investment, together with attributable research, training and industrial infrastructure |

B2AGI applies these five lenses to compare industrialization pathways. The assessment combines documented activities with quantitative indicators in their original units and periods. It uses qualitative comparison rather than a composite numerical score.

## Geographic Order

The assessment prioritizes Intelligence and Beyond: the capacity to develop intelligent physical systems and the investment in knowledge, people and industrial infrastructure that carries that capacity forward. Physical applications, capital access and operating responsibility establish how those capabilities reach practical work. Robot production and installation volume are supporting indicators.

Industrial-robot installations, robot density and R&D expenditure were examined alongside operating reports, technical research and institutional records. Each statistic retains its observation year, unit and geographic definition. The source references attached to each entry let readers examine the basis of the assessment.

### Reading the Leading Countries

The United States leads through computing and robotics platforms that other developers use, university research and shared development tools. China follows through sensing and robot development connected to electronics engineering, manufacturing and shared training facilities. The comparison recognizes different strengths: the US profile has broader reusable development platforms; China's profile connects component engineering to product development and industrial demand.

South Korea's case combines investment in research with several distinct technical contributions: NAVER's cloud-based robot coordination, KAIST's learned locomotion control and university-led technology commercialization. WIPO's 2026 human-capital-and-research pillar places Korea first, Germany third, Sweden fifth, Japan eighteenth and Israel nineteenth. These economy-wide indicators provide context for the robotics cases. Korea's 2024 R&D expenditure was 5.13% of GDP, compared with Japan's 3.62% and Germany's 3.13%.

### Country Comparison Record

Comparison date: 11 October 2026. Intelligence and Beyond remain the assessment priorities. The evidence concerns technical contributions, investment and the institutions developing future capabilities.

| Country | Intelligence contribution | Beyond contribution | Strongest competing case |
| --- | --- | --- | --- |
| South Korea | NAVER's ARC robot coordination; KAIST's APT-RL controller tested on HOUND indoors and outdoors | WIPO research context; KAIST's robotics alliance reported technology-transfer and investment activity in its first year | NAVER's building is an internal environment. KAIST's March 2026 testbed demonstration and its plans for open access describe different stages. |
| Japan | Mujin's perception and motion generation at SUBARU's Oizumi factory | ABCI 3.0 shared computing with an application and allocation process | Existing computing access and factory engineering provide development capability beyond robot production. |
| Germany | Sereact's vision-language-action manipulation and Fraunhofer IEM's sensor-guided development | Partner-access equipment, prototype services and robotics training | Direct robotics development services strengthen both Intelligence and Beyond. |
| Sweden | WARA Robotics' navigation-and-manipulation challenge linked to AstraZeneca | WASP doctoral training and the ABB-hosted research arena | Sweden exceeds Korea's 2021 investment reference and has a documented industry-linked research and training programme. |
| Israel | Mentee's observation-based manipulation development | National AI computing reported operational in January 2026, with subsidized allocations | Israel exceeds Korea's R&D intensity and 2021 investment reference. The computing programme supports academia and industry. |

South Korea ranks third for the combination of operating robot coordination, physically tested learned control and university-led robotics commercialization, supported by its research and human-capital base. NAVER's building operations, KAIST's HOUND experiments and the robotics alliance's reported transfer and investment activity contribute separately to that assessment. Japan's manufacturing intelligence, Germany's manipulation and prototyping services, Sweden's research and doctoral training, and Israel's manipulation development and computing access provide the competing profiles. Korea's combined profile carries the comparison; Sweden and Israel's higher economy-wide investment references remain important context.

Japan ranks fourth and Germany fifth. Mujin's SUBARU application connects perception and motion generation to an established manufacturing process, while ABCI provides shared computing through its application and allocation process. Germany contributes Sereact's vision-language-action manipulation and Fraunhofer IEM's partner-access development, prototyping and training. In this close comparison, Japan's factory-process engineering together with shared development infrastructure takes the higher place. Both profiles contain deployed intelligence and resources for further research.

Sources: [KAIST HOUND](https://www.kaist.ac.kr/newsen/html/news/?mode=V&mng_no=64450), 2026-07-16, APT-RL and physical experiments; [KAIST commercialization programme](https://www.kaist.ac.kr/newsen/html/news/?mode=V&mng_no=58790), 2026-02-27, alliance and first-year results; [NAVER 1784](https://navercorp.com/story/storyDetail?seq=32450), 2025-04-03; [Mujin/SUBARU](https://www.mujin.co.jp/news/9045/), 2024-05-15; [ABCI access](https://abci.ai/en/how_to_use/), FY2026 conditions; [Sereact](https://sereact.ai/posts/sereact-fundraising-series-a), 2025-01-20; [Fraunhofer IEM](https://www.iem.fraunhofer.de/de/ueber-uns/labore-pruefeinrichtungen/robotics-lab.html), laboratory services; [WARA challenge](https://wasp-sweden.org/phd-students-challenged-by-the-industry-at-wara-robotics-challenge/), 2025-01-21, December 2024 event; [WASP graduate school](https://wasp-sweden.org/graduate-school/), May 2026 graduate statistic; [Israel computing access](https://innovationisrael.org.il/en/press_release/supercomputer-access-2026/), 2026-01-18. WIPO's research ranks and observation periods are in Country Context.

### Country Context

The same research and institutional indicators were examined alongside the robotics records. WIPO's human-capital-and-research and institutions ranks describe the general innovation environment. They provide context for the sector evidence. Taiwan's R&D observation comes from NSTC; the other observations use the WIPO database.

| B2AGI rank | Economy | R&D expenditure, % GDP | Observation year | WIPO research rank | WIPO institutions rank |
| ---: | --- | ---: | ---: | ---: | ---: |
| 1 | United States | 3.44 | 2024 | 14 | 19 |
| 2 | China | 2.69 | 2024 | 23 | 48 |
| 3 | South Korea | 5.13 | 2024 | 1 | 22 |
| 4 | Japan | 3.62 | 2024 | 18 | 23 |
| 5 | Germany | 3.13 | 2024 | 3 | 21 |
| 6 | United Kingdom | 2.75 | 2023 | 7 | 28 |
| 7 | Switzerland | 3.10 | 2023 | 8 | 5 |
| 8 | France | 2.18 | 2024 | 17 | 35 |
| 9 | Canada | 1.81 | 2025 | 11 | 15 |
| 10 | Taiwan | 4.08 | 2024 | - | - |
| 11 | Denmark | 3.01 | 2024 | 12 | 2 |
| 12 | Sweden | 3.56 | 2024 | 5 | 11 |
| 13 | Singapore | 2.03 | 2023 | 2 | 1 |
| 14 | Netherlands | 2.29 | 2024 | 16 | 13 |
| 15 | Israel | 6.76 | 2024 | 19 | 33 |
| 16 | Australia | 1.68 | 2023 | 9 | 12 |
| 17 | Italy | 1.38 | 2024 | 29 | 53 |
| 18 | Spain | 1.50 | 2024 | 27 | 55 |
| 19 | Norway | 1.80 | 2024 | 22 | 8 |
| 20 | United Arab Emirates | 1.49 | 2021 | 13 | 9 |

Sources: [WIPO GII 2026 database](https://www.wipo.int/edocs/pubdocs/en/wipo-pub-2000-2026-tech1.xlsx), indicators 2.3.2, 2 and 1; [indicator definitions](https://www.wipo.int/web-publications/global-innovation-index-2026/en/appendix-iii-sources-and-definitions.html); [Taiwan NSTC statistics](https://wsts.nstc.gov.tw/stsweb/technology/TechnologyStatistics.aspx?ID=1&language=E). WIPO ranks cover 139 economies and use each index's reporting periods; Taiwan is outside that database. The UAE observation is from 2021. Industrial demand uses the [IFR World Robotics 2026 release](https://ifr.org/ifr-press-releases/news/five-million-robots-now-operate-in-factories-globally); Japan's manufacturing supply uses [JARA calendar-2025 results](https://www.jara.jp/e/data/dl/year/IR_CY2025_e.pdf). The country records provide the sector-specific sources and comparisons.


### Beyond: Investment in Future Capacity

Beyond examines investment in knowledge, people and the industrial foundation for the next generation of physical intelligence. The country assessment uses R&D expenditure, government education expenditure and gross fixed capital formation, each expressed as a share of GDP. Regional assessment uses attributable research, training and development infrastructure within each cluster's boundary.

The reference calculation covers the original 24 candidate economies. Within the 2019-2023 observation window, 2021 provides complete three-indicator coverage for 19 economies, the largest coverage and the latest year at that coverage level. The reference equals the mean of three normalized ratios: each observed value divided by the highest observed value among the 24 candidates, multiplied by 100. The 2021 denominators are 5.75725% for R&D, 7.60224008560181% for education and 41.3521203793076% for fixed capital. A dash records a missing observation; a reference is calculated when all three observations are present.

The table reports the preserved 2021 WDI snapshot. It describes economy-wide investment intensity. The country assessment reads that investment alongside the attributed intelligence, products, customer work and infrastructure in each profile. Finland, India, Saudi Arabia and Turkey are comparison candidates outside the selected 20.

The Beyond reference is the equal-weight mean of three normalized, economy-wide expenditure shares. It provides context for the qualitative assessment; the B2AGI order comes from the comparative assessment of the country profiles. Government education expenditure and gross fixed capital formation describe broad investment in future capacity. Robotics training, venture funding and access to shared robotics facilities are assessed through their own attributed records.

The 2024 R&D observations in Country Context update one indicator. The 2021 reference combines three indicators observed in the same year. Substituting a 2024 R&D value while retaining 2021 education and capital values would produce a mixed-year calculation. A same-year update requires all three series for the candidate economies and new normalization denominators.

| Economy | R&D, % GDP | Education, % GDP | Fixed capital, % GDP | Beyond reference |
| --- | ---: | ---: | ---: | ---: |
| United States | 3.47 | 5.42 | 21.29 | 61.01 |
| China | 2.38 | - | 41.35 | - |
| South Korea | 4.60 | 4.56 | 31.49 | 71.97 |
| Japan | 3.27 | 3.34 | 27.08 | 55.41 |
| Germany | 3.08 | 5.37 | 21.18 | 58.43 |
| United Kingdom | 2.90 | 5.91 | 18.54 | 57.63 |
| Switzerland | 3.25 | 5.03 | 26.78 | 62.43 |
| France | 2.21 | 5.42 | 23.48 | 55.49 |
| Canada | 1.87 | 4.75 | 23.79 | 50.83 |
| Taiwan | - | - | - | - |
| Denmark | 2.74 | 6.96 | 23.21 | 65.10 |
| Sweden | 3.42 | 7.60 | 25.12 | 73.35 |
| Singapore | 1.94 | 2.83 | 22.08 | 41.41 |
| Netherlands | 2.22 | 5.38 | 20.69 | 53.08 |
| Israel | 5.76 | 6.17 | 24.04 | 79.75 |
| Australia | 1.86 | 5.34 | 22.58 | 52.38 |
| Italy | 1.41 | 4.17 | 20.77 | 43.20 |
| Spain | 1.40 | 4.81 | 20.20 | 45.47 |
| Norway | 1.89 | 6.96 | 24.00 | 60.79 |
| United Arab Emirates | 1.49 | 3.89 | - | - |
| Finland | 3.01 | 6.59 | 24.42 | 66.01 |
| India | - | 4.63 | 29.58 | - |
| Saudi Arabia | 0.39 | - | 23.52 | - |
| Turkey | 1.40 | 3.54 | 28.39 | 46.52 |

Sources: World Bank WDI [R&D expenditure](https://data.worldbank.org/indicator/GB.XPD.RSDV.GD.ZS), [government education expenditure](https://data.worldbank.org/indicator/SE.XPD.TOTL.GD.ZS) and [gross fixed capital formation](https://data.worldbank.org/indicator/NE.GDI.FTOT.ZS). Indicator IDs: `GB.XPD.RSDV.GD.ZS`, `SE.XPD.TOTL.GD.ZS`, `NE.GDI.FTOT.ZS`. Values and the reference are rounded to two decimals for display; the calculation uses the original precision.

Preserved observations: [R&D JSON](sources/wdi/GB.XPD.RSDV.GD.ZS.json), [education JSON](sources/wdi/SE.XPD.TOTL.GD.ZS.json), [fixed-capital JSON](sources/wdi/NE.GDI.FTOT.ZS.json). These WDI snapshots accompany the reference calculation. World Bank publishes the indicators under CC BY 4.0.

### Reading the Leading Clusters

Shenzhen-Dongguan connects sensing and robot engineering with university-linked product development. XbotPark's incubation history names Narwal and Hai Robotics. An April 2026 HKUST student visit documents continuing education and industry contact at its Dongguan base. Shenzhen's report on 2024 activity also records use of the Nanshan entrepreneurship academy and 16 incubated angel-stage projects. That academy total covers hardware entrepreneurship, rather than 16 robotics companies.

The Bay Area combines reusable robotics platforms with Berkeley's robot-learning curriculum and Stanford's shared physical research facilities. Stanford Robotics Center opened in November 2024 and includes experimental space for logistics, manufacturing and medical robotics. Physical experimentation and engineering education belong to both regional profiles.

### Cluster Comparison Record

| Question | Shenzhen-Dongguan | San Francisco Bay Area |
| --- | --- | --- |
| What intelligence is developed? | Located sensing and robot products, including RoboSense's perception hardware | Model, simulation and robot-learning platforms, including NVIDIA's developer tools |
| How are people and future capabilities developed? | University-linked engineering, incubation and supplier-supported prototyping | Berkeley's dated learning curriculum and Stanford's collaborative experimental facilities |
| What activity connects these resources to use? | Named incubation outcomes; HKUST's April 2026 educational visit; Shenzhen's reported academy incubation | Named platform adopters; published course work; Stanford's task-focused research bays |
| What remains distinct? | Incubation totals include non-robotics hardware; Hong Kong educational origins and overseas customers retain their locations | Global platform adoption is distinct from local use; facility descriptions are distinct from completed transfer outcomes |

Shenzhen-Dongguan ranks first for the documented connection between university-linked engineering education, incubation and local robot product development with manufacturing access. XbotPark's named incubation outcomes give this pathway concrete company examples; the April 2026 visit documents continuing educational contact. The Bay Area ranks second with reusable robotics tools, Berkeley's learning curriculum and Stanford's shared physical research facilities. Both develop intelligence and future engineering capacity. The local education-and-incubation connection to robot engineering carries the leading-place decision. Incubation records, educational visits and product outcomes retain their distinct dates and scopes in the sources.

Sources: [HKUST incubation account](https://www.hkust-gz.edu.cn/2022/10/15/li-zexiang-and-ko-ping-keung-talk-about-innovation-and-entrepreneurship-at-hkust-gz/), 2022-10-15, named startup outcomes; [HKUST educational visit](https://ec-gba.hkust.edu.hk/news/gba-entrepreneurship-learning-trip-2d1n-journey-30-aspiring-minds), 2026-04-24, Dongguan visit; [Shenzhen government report](https://www.sz.gov.cn/zfgb/2025/1366/content/post_12090037.html), published 2025-03-24, section on 2024 education and research activity; [Stanford Robotics Center](https://src.stanford.edu/the-center), research bays and 2024-11-01 launch. Existing developer and course references accompany cluster:01 and cluster:02.

The Yangtze River Delta profile combines Shanghai's training infrastructure with Hangzhou's robots and physical-system research. Beijing brings together learning resources, developers, local autonomous-driving tests and funding. The Seoul Capital Area connects research, manufacturing and logistics demand, office deliveries and capital-market access. These are different industrial pathways, assessed as defined regions rather than equal-sized administrative units.

Cluster boundaries and the footprint of cited evidence are recorded separately. A three-city analytical boundary does not establish activity in every city. Customer sites outside the boundary retain their own locations.

Training sources: [XbotPark Robotics Academy](https://www.xbotpark.com/en/TalentDevelopment/), [XbotPark professional training](https://www.xbotpark.com/en/TrainingPrograms/) and [UC Berkeley CS 185/285, Spring 2026](https://rail.eecs.berkeley.edu/deeprlcourse/). The academy account dates its establishment to 2015; its programme descriptions identify training and facilities, while Berkeley publishes a dated 2026 course. Each retains its own reporting period.

| Rank | Cluster | Record |
| ---: | --- | --- |
| 1 | Pearl River Delta Core — Shenzhen–Dongguan | cluster:01 |
| 2 | San Francisco Bay Area | cluster:02 |
| 3 | Yangtze River Delta Core — Shanghai–Hangzhou–Suzhou | cluster:03 |
| 4 | Beijing Region | cluster:04 |
| 5 | Seoul Capital Area | cluster:05 |
| 6 | Boston Region | cluster:06 |
| 7 | Pittsburgh Region | cluster:07 |
| 8 | Munich Region | cluster:08 |
| 9 | Tokyo–Yokohama Core | cluster:09 |
| 10 | Stuttgart Region | cluster:10 |
| 11 | Zurich Region | cluster:11 |
| 12 | Singapore | cluster:12 |
| 13 | Odense | cluster:13 |
| 14 | Seattle Region | cluster:14 |
| 15 | Aichi Manufacturing Region | cluster:15 |
| 16 | London Core — Cambridge Research and Startup Lineage | cluster:16 |
| 17 | Waterloo Core | cluster:17 |
| 18 | Daejeon Core — Sejong–Osong Transport Links | cluster:18 |
| 19 | Pohang Industrial Robotics and Manufacturing Research Region | cluster:19 |
| 20 | Austin Region | cluster:20 |

## Role-Based Groups

**Current 100** recognizes established capabilities adopted in production, operations or education and research, alongside completed technology transfers and capital contributions. **Next 100** follows new robot, policy and application programmes in development, prototype testing or research validation. Each entry names the contribution being assessed. A supplied arm can belong to Current while the new manipulation policy tested on it belongs to Next.

**Influence 50** recognizes attributed technical methods, robot designs, task-specific research leadership and operating or strategic decisions. **Ecosystem Builders 50** recognizes shared tools, education, standards, finance and research or testing facilities. Each entry names the contribution that determines its group; collaborative results retain shared attribution.

### People: Contribution Decisions

The following examples show how the contribution determines the group.

| Records | Contribution recognized | Result |
| --- | --- | --- |
| E019 Oh Jun Ho; P038 In So Kweon | KAIST identifies both as leaders of the DRC-HUBO team | Both are Influence selections |
| E006 Chelsea Finn; P015 Karol Hausman | The pi0.5 contribution table names Finn under policy training and research, and Hausman under robot infrastructure and writing | Finn: Influence; Hausman: Ecosystem Builders |
| P010 Gill Pratt | DARPA records his direction of a common test programme for competing teams | Ecosystem Builders |
| P051 Nikita Rudin; E022 Russ Tedrake | Rudin's selected work is a locomotion-training method; Tedrake's is the shared Drake software project | Rudin: Influence; Tedrake: Ecosystem Builders |
| P006 Daniela Rus | Completed doctoral mentoring and support for a researcher's development | Ecosystem Builders |
| P036 Jiyang Gao | Joint supervision of Galaxea's dataset-and-model programme; this entry selects coordination of shared research capability | Ecosystem Builders |
| E033 Wang Tianran; E034 Zhao Mingguo | Attributed industrial-robot product decisions and joint humanoid-soccer research leadership | Influence |

For Gao, the selected contribution is coordination of a shared research programme. Oh and Zhao's entries select leadership of specific robot-team results. Galaxea's paper credits dataset operators and policy trainers separately, and its first version describes public release as forthcoming. Research supervision can support either group; the programme or technical contribution chosen for the entry determines the assignment.

Institutional entries identify research, testing, regulatory, industry-support and capital-market activities. Regulators and exchanges are assessed for their contribution to operating conditions and capital access.

For worker-safety coverage, I054 NIOSH contributes fatality research and joint analysis of Ohio injury claims alongside I036 OSHA's inspector-facing guidance. A3's engineering contribution remains credited in the joint guidance. NIOSH's empirical hazard research adds evidence of observed worker harm to standards and guidance.

The 350 role-based entries are selected within fixed category sizes, with a Korea-China-US research focus. Company selection is separate from references to that company in geographic or individual profiles.

### Reading Current and Next in Practice

| Record | Selected activity | Group choice |
| --- | --- | --- |
| C044 Gecko Robotics | Robotic inspection and AI analysis in BPMI component trials | Next: the selected inspection application is under validation |
| C088 Siemens | Control and integration in Erlangen factory tote-handling trials | Next: the selected integration is a developing factory application |
| C105 Doosan Robotics | Arm supplied to an external vision-guided welding project | Current: an adopted manipulation platform supports the team's new application |
| C046 GelSight | Mini tactile sensor used by external researchers for robot-mounted data acquisition | Current: adopted instrumentation; the researchers retain their inspection methods |
| N072 Shadow Robot | Released DEX-EE hand used in DeepMind manipulation research | Current: jointly developed and adopted research hardware |
| N111 Hello Robot | Stretch AI language-directed navigation and manipulation integration | Next: the company's developing software application |
| N112 Wayve | LINGO-2 trajectory-and-commentary model and simulated language-conditioned manoeuvres | Next: the named 2024 driving research contribution |
| N113 Hugging Face | SmolVLA perception-to-action model and physical manipulation evaluations | Next: the named research programme; I026 separately recognizes adopted LeRobot infrastructure |
| N020 Farm-ng | Amiga used in the recurring Farm Robotics Challenge | Current: an adopted education and development platform |
| N034 Intel | Loihi 2 computing in joint Astrobee control research | Next: the selected robot-control programme is feasibility research |
| N003 Aigen | Element's initial use and evaluation at Bowles farm | Next follows the emerging agricultural application |
| N008 Apptronik | Apollo transport and part-handling trials at Mercedes-Benz | Next follows the humanoid product's customer-development pathway |
| N056 Outrider | Customer-site movement and learned connection trials | Next follows the emerging autonomous-yard application |

The groups classify the selected contribution. An instrument supplier retains its hardware contribution when an external team develops a new method using that instrument. Evidence scope specifies customer trials, acceptance, recurring operation and reported performance at the record level.

Hello Robot is selected for Stretch AI's integration of semantic memory, navigation and manipulation. Wayve is selected for LINGO-2's dated language-conditioned driving research, with road and simulation results distinguished. Rainbow Robotics and DeepRoute.ai are comparison candidates outside this edition's company selection. Rainbow's research-platform supply and DeepRoute's production-assistance and learning work are the alternative cases documented in [Selection Comparisons](#selection-comparisons). Momenta and Horizon provide other Chinese automotive-software and computing coverage.

Waymo contributes documented driverless passenger operation; UBTECH contributes humanoid production and delivery linked to factory tasks. DJI supplies infrastructure-inspection capabilities, and Doosan supplies the arm identified in the external project team's engineering account. Toyota/TRI, Torc, Rhoda AI and RIVR contribute whole-body manipulation, highway freight, video-pretrained industrial handling and terrain-aware delivery programmes. The edition includes 100 Current and 100 Next companies. Membership and group changes are listed in the [revision history](CHANGELOG.md).

## Selection Comparisons

Decision date: 2026-10-11. Each comparison identifies the selected activity, the alternative's strongest evidence and the reason for allocating the place. Current and Next describe the selected contribution's stage. Several operating suppliers give way to research programmes that add a different technical question to the edition.

### Perception-to-Action Research: Hugging Face / RealSense

N113 Hugging Face is selected for SmolVLA's compact vision-language-action model, community-data training and physical manipulation experiments. N063 RealSense contributes adopted RGB-D camera input to LimX's navigation demonstration; NVIDIA supplies the odometry and LimX integrates navigation. That camera contribution belongs with established supply. The emerging-research place goes to SmolVLA for its inspectable model experiments and task-adaptation resources.

RealSense retains supplier credit and its source reference under N042 LimX. C068 Orbbec supplies other depth-sensing coverage. The selection gives up RealSense's standalone corporate profile. SmolVLA adds a compact, community-data approach alongside the existing model researchers; joint authors, hardware designers and data contributors retain credit. I026's LeRobot entry concerns shared infrastructure, while the company entry concerns the developing model.

Sources: [SmolVLA paper](https://arxiv.org/html/2506.01844v1), June 2025, author affiliations and physical experiments; [Hugging Face research account](https://huggingface.co/blog/smolvla), 2025-06-03, architecture and SO100/SO101 evaluations; [model card](https://huggingface.co/lerobot/smolvla_base), intended use and adaptation; [RealSense navigation account](https://www.realsenseai.com/news-insights/insights/advancing-autonomous-humanoid-navigation-with-realsense-depth-cameras-and-nvidia-cuvslam/), camera contribution and LimX integration.

### Driverless Passenger Service: Waymo / May Mobility

C102 Waymo is selected for its documented rider-only service across five regions through June 2026 and published deployment conditions. C059 May Mobility's Peachtree Corners service is a concrete commercial driver-out operation with eight stops. The place goes to Waymo for the broader documented service footprint. May's service retains its operating status. Waymo's manufacturer recall is part of its operating evidence, not erased by selection.

Sources: [Waymo Safety Impact](https://www.waymo.com/safety/impact/), rider-only miles and city table through June 2026; [CPUC permit table](https://www.cpuc.ca.gov/regulatory-services/licensing/transportation-licensing-and-analysis-branch/autonomous-vehicle-programs/autonomous-vehicle-program-permits-issued), Waymo driverless deployment; [May Mobility](https://maymobility.com/posts/may-mobility-now-offers-autonomous-driverless-rides-in-peachtree-corners/), 2025-02-12, launch and service details; [Waymo recall](https://static.nhtsa.gov/odi/rcl/2025/RCLRPT-25E034-2471.PDF), 2025-05-12, pp. 1-3.

### Industrial Humanoids: UBTECH / Kepler

C103 UBTECH is selected for dated production and delivery disclosure linked to factory handling, sorting and inspection. N038 Kepler's K2 training at SAIC-GM supplies a named factory-development case. UBTECH's production and delivery milestone determines the choice for this supply contribution; Kepler's factory access is the strongest competing evidence. The cited UBTECH solution page describes NIO, BYD, Geely Auto, Dongfeng Liuzhou Motor and FAW-Volkswagen Qingdao separately, including training relationships.

Sources: [UBTECH HKEX results](https://www.hkexnews.hk/listedco/listconews/sehk/2026/0331/2026033102607.pdf), filed 2026-03-31, printed pp. 2-4 and 10-11; [industrial solutions](https://www.ubtrobot.com/en/humanoid/solutions/industry), manufacturer-specific sections, undated; [Kepler issuer release](https://www.prnasia.com/story/486457-1.shtml), 2025-04-23, K2 training at SAIC-GM.

### Aerial Inspection: DJI / Skydio

C104 DJI is selected for the named China Southern Power Grid inspection workflow, accompanied by separately dated Dock development evidence. C089 Skydio's Caltrans project documents off-grid dock deployment and a public-agency validation pathway. The completed historical inspection task favors DJI for this place; Skydio contributes a specific public-infrastructure project that this substitution leaves outside the list. The 2019 task and later Dock functions retain separate dates.

Sources: [DJI power-grid case](https://enterprise-insights.dji.com/user-stories/china-southern-grid-powerline-inspection), 2019-05-08, Intelligent Machines at Work; [DJI Dock workflows](https://enterprise-insights.dji.com/blog/important-takeaways-from-automating-oil-gas-inspections-with-dji-dock), 2026-02-20; [Skydio Caltrans case](https://www.skydio.com/customer-stories/caltrans-solar-powered-dock), Synopsis, undated; [Caltrans project sheet](https://www.transportation.gov/sites/dot.gov/files/2024-11/Caltrans_508.pdf), Stage 1 Outcomes and Stage 2 Vision.

### Supplied Manipulation Platforms: Doosan / Dobot

C105 Doosan is selected for the arm identified in the Royal IHC project team's dated welding-prototype account. C031 Dobot's Bristol case documents MG400/CR3 supply for tactile and Braille research. The dated project-author description of hardware, vision and control provides the deciding evidence resolution. Dobot's supplied research platform remains a substantive alternative; its Braille application is coverage lost in this choice. The welding method belongs to the external team.

Sources: [Royal IHC project account](https://www.robotminor.nl/pub/versatile-welding-cobot-royal-ihc), published 2025-03-16, fall 2023 project, Solution, Integration and Program; [Dobot Bristol case](https://www.dobot-robots.com/insights/case-studies/driving-innovation-at-bristol-robotics-laboratory.html), MG400/CR3 and Ao Li's research account, undated.

### Manipulation Research: Toyota/TRI / Dexmate

N107 Toyota/TRI adds the jointly developed language-conditioned whole-body policy and physical evaluation. C028 Dexmate documents Vega adoption by named research users. With supplied research platforms also represented by Unitree and other selected companies, this place adds coordinated locomotion and manipulation learning. The substitution loses Vega-specific adoption coverage; Toyota/TRI's contribution retains Boston Dynamics' joint credit.

Sources: [joint Atlas research](https://bostondynamics.com/blog/large-behavior-models-atlas-find-new-footing/), policy pipeline and Long-Horizon/Approach sections, undated; [Dexmate](https://www.dexmate.com/), Vega hardware and named-user descriptions, undated.

### Freight: Torc / Gatik

N108 Torc adds driver-out highway-system validation toward a stated 2027 commercial target. C042 Gatik has stronger completed-operation evidence: Walmart confirms daily driverless routes. Retained Kodiak supplies an operating freight pathway; Torc adds a bounded highway validation programme. This allocation leaves Gatik's particular retail middle-mile route outside the edition and does not convert Torc's closed-course tests into public-road driverless service.

Sources: [Torc validation](https://torc.ai/wp-content/uploads/2024/10/Torc-Robotics-Performs-Successful-Fully-Autonomous-Product-Validation.pdf), 2024-10-29; [Torc FAQ](https://torc.ai/frequently-asked-questions/), testing conditions and commercialization timing; [Walmart/Gatik](https://corporate.walmart.com/news/2020/12/15/walmart-and-gatik-go-driverless-in-arkansas-and-expand-self-driving-car-pilot-to-a-second-location), Editor's Note updated 2021-11-08.

### Industrial Handling: Rhoda / Mech-Mind

N109 Rhoda adds a defined video-pretraining experiment with physical bearing-unpacking trials. C060 Mech-Mind documents camera-guided picking, scanning and shelf transfer at an unnamed FMCG warehouse. Covariant and OSARO retain implemented picking coverage; Rhoda adds a model-to-task evaluation with explicit trial scoring. Mech-Mind's camera and truss implementation is lost coverage. Rhoda's single-setup experiment has its stated research scope.

Sources: [Rhoda research](https://www.rhoda.ai/research/scaling-web-video-pretraining), 2026-09-10, task, scoring, evaluation and Limitations; [Mech-Mind case](https://www.mech-mind.com/cases/vision-guided-case-goods-sorting), Customer Demand and Solution, undated.

### Ground Delivery: RIVR / Kiwibot

N110 RIVR adds Veho's development experiment on stairs, gates and uneven approaches. C056 Kiwibot has stronger service evidence in Sodexo's supported campus deliveries. Meituan and Vayu preserve other ground-delivery pathways; RIVR adds terrain access and courier handoff as development questions. Campus food and point-of-sale integration are lost coverage. Veho's later launch account stays separate from the initial experiment's task conditions.

Sources: [Veho experiment](https://www.shipveho.com/blog/ai-meets-logistics-veho-and-rivrs-strategic-partnership-in-robotic-parcel-delivery), 2025-05-27, What We're Testing; [Veho 2026 account](https://www.shipveho.com/news/veho-named-to-fast-company-most-innovative-2026), RIVR launch paragraph; [Sodexo/Kiwibot](https://www.sodexo.com/news/newsroom/2022/kiwibot-roll-out-USA), February 2022, existing operations and support.

### Mobile Manipulation: Hello Robot / Rainbow Robotics

N111 Hello Robot adds Stretch AI's company-maintained semantic-memory, navigation and manipulation integration. C077 Rainbow Robotics supplies RB-Y1 for external robot-learning research. That adopted platform is a strong Current contribution. The place is allocated to Hello Robot's developing integration, while Unitree, ROBOTIS and other entries retain supplied-platform coverage. Rainbow-specific mobile bimanual hardware adoption is lost coverage. Hello Robot's source credits inherited research methods separately.

Sources: [Stretch AI](https://github.com/hello-robot/stretch_ai), README, Hardware Requirements and Language-Directed Pick and Place; [OVMM](https://github.com/hello-robot/stretch_ai/blob/main/docs/ovmm.md), Pipeline; [Ai2 MolmoBot](https://allenai.org/blog/molmobot-robot-manipulation), 2026-03-11, updated 2026-03-27, One suite, many tasks and How it performs: RB-Y1 physical manipulation and transfer. Repository descriptions were reviewed on 2026-10-11.

### Learned Driving: Wayve / DeepRoute.ai

N112 Wayve adds LINGO-2's joint trajectory/commentary model and language-conditioned simulation experiments. C027 DeepRoute.ai supplies learned production-assistance systems and describes its own 40B VLA architecture and training-data pipeline. Momenta and Horizon retain other Chinese automotive-software and computing contributions. This place adds the explicit explanation/action research question; it loses DeepRoute's particular VLA and production pathway. LINGO-2's 2024 research is the selected contribution, distinct from Wayve's later services.

Sources: [LINGO-2](https://wayve.ai/thinking/lingo-2-driving-with-language/), 2024-04-17, Architecture, Ghost Gym and Limitations; [DeepRoute IO 2.0](https://www.prnewswire.com/news-releases/deeprouteai-brings-mass-production-ready-deeproute-io-2-0-platform-to-iaa-2025--302549100.html), 2025-09-08; [DeepRoute 40B VLA](https://www.prnewswire.com/news-releases/deeprouteai-presents-40b-vision-language-action-foundation-model-at-nvidia-gtc-2026-accelerating-autonomous-driving-at-scale-302716046.html), 2026-03-17, One Model, Three Roles and A Self-Evolving Data Flywheel.

### Worker-Safety Evidence: NIOSH / A3

I054 NIOSH adds fatality research and joint analysis of injury claims alongside OSHA's guidance. A3 contributes engineering expertise to that shared guidance and has distinct standards work. The place goes to empirical worker-harm evidence to complement retained standards and guidance coverage. I002 (A3) was retired on 2026-10-11; its joint-guidance credit is retained in I054. The I054 source references identify the joint guidance and the two studies with their observation periods.

## Evidence and Geographic Attribution

Evidence is classified by activity: research, simulation, physical demonstration, supervised trial, customer operation, financing or certification. Plans, signed agreements and completed transactions retain their respective status and dates.

Geographic attribution follows the activity:

- A customer operation belongs to the location of the customer site.
- Research, development and manufacturing belong to their respective locations.
- Corporate and financing contributions are recorded separately from customer deployment.
- Cross-references identify participants in shared events. Repeated citations do not establish additional customer activity.

For example, [ABB's 2020 announcement](https://new.abb.com/news/detail/57457/abb-and-covariant-partner-to-deploy-integrated-ai-robotic-solutions) places the Active Ants installation in Utrecht. The Netherlands profile's DLL reference concerns financing for a project in New Zealand. [Amazon's robotics article](https://www.aboutamazon.com/news/operations/how-amazon-deploys-robots-in-its-operations-facilities) places trials near Nashville and in Houston, and describes sorting in San Marcos. The Seattle profile recognizes the corporate contribution separately from those customer-site activities. Amazon and Roboto have separate source records.

Source records identify the publisher and URL. Reported outcomes remain attributed to the company, customer or institution making the report. Test conditions and human supervision stay attached to the task described.

## Statistics and Financial Data

Statistics retain the observation year, release date, unit, denominator and geographic definition. Industrial-robot installations measure manufacturing automation demand; humanoid shipments and autonomy require their own indicators.

Financial values require the reporting entity, currency, scale, fiscal period or balance-sheet date and accounting basis. Revenue and operating income are period measures; cash, assets and liabilities are date-specific balances. Consolidated company results describe the company, not an inferred robotics segment.

## Review of Contested Decisions

Comparisons record the deciding evidence, the strongest competing case and the resulting decision. Rank changes, selection changes and factual corrections remain separately identifiable.

### Country and Cluster Positions

A contested position requires a comparison with the nearest relevant alternatives through the same five lenses. The decision record must identify the evidence that determines the order, the strongest evidence favoring the alternative, the geographic boundaries and reporting periods, and the reasons for the resulting placement. Sparse documentation must be distinguished from weak industrial capability. Historical contributions and current operations must retain their respective dates.

### Company and Institution Selection

The company and institution lists are fixed-size contribution selections. Published selection comparisons identify the candidate, the contribution considered, comparable entries, the applicable scope and the reason for the choice.

Institutional comparisons must distinguish technical authorship, research, funding, standards administration, regulatory action and operating support. When several participants share an output, each retains its contribution. The selection decision must explain the allocation of places; a shared source does not supply that explanation.

### People and Group Assignment

A disputed assignment must be tested against the contribution described in the record. The review compares the person's work with relevant entries in both groups and explains which contribution determines the assignment. Group names are not a ranking of prestige. Changes preserve the record's identity and acknowledge collaborators.

### Evidence Submission and Decision Records

An objection should identify the affected record or omitted subject, the disputed claim, relevant comparator records and dated primary sources with an exact page or section. A factual correction, a group change and a rank or selection change require separate decisions.

Each reassessment must record the evidence considered, the result, the reason and its date. New comparisons use the date on which they are performed. The published revision history records changes to ranks, selections and groups as well as factual corrections.

## Citation and Corrections

Use this format:

> B2AGI Physical AI Ranking 2027, record [ID], reference [reference_id], revision [revision or release identifier]. Original source: [publisher], [title], [publication date], [URL], [page or section].

The CSV's `source_references` and `source_urls` contain paired, newline-separated entries in the same order. For example, C044's first reference is `C044:S1`; its URL leads to Gecko's BPMI inspection account. The JSON stores that pair in `sources`, together with the publisher domain, original `evidence_id` and recorded access date. The workbook uses the same references in its Sources sheet.

Use the public `reference_id` to cite any source. The `evidence_id` field carries its matched identifier from the original evidence registry. Exact URL matches reuse an existing registry ID; ambiguous matches retain the public reference alone.

Record IDs remain stable when the display order changes. Gaps in their numbers preserve record identity; category totals are calculated from the records present. Spreadsheet row numbers help navigation and can change between exports.

The JSON's `recorded_access_date` stores the source-reading date where recorded. Publication and event dates are available in the cited publications and, where relevant, in the rationale or evidence scope. Empty date fields remain blank.

A correction identifies the affected record, previous claim, replacement claim, supporting source and correction date. Changes to the selection set or rank are recorded separately from factual or wording corrections. Each published release supplies its date and version identifier so readers can cite the copy they used.

Send corrections and selection evidence to [aei.b2agi@gmail.com](mailto:aei.b2agi@gmail.com), including the record ID, source URL and relevant page or section.

Assessment baseline: October 2026. English edition: v1.0, 2026-10-11. [Revision history](CHANGELOG.md).