Who counts, and how
Search, compare and verify specifications field by field. Missing information is never estimated.
- quadruped robots
- 167
- humanoid robots
- 365
- robot manufacturers
- 260
- countries
- 23
- robot AI models
- 94
- AI model developers
- 57
297 companies in all; 20 make both robots and AI models.
Figures as of 28 October 2026
How a figure reaches the page
-
Fetched
We fetch the manufacturer’s own page, datasheet or manual and keep the file as evidence, with its address and the date it was fetched.
-
Recorded
The figure goes into one of the schema’s 33 fields with a unit and a source.
-
Checked for completeness
The validator reads the whole catalogue again and stops the build if a figure is missing a unit, a source or a date.
-
Built
The pages are generated afresh from the same files, and the build runs the validator first every time.
Everything the catalogue can know
Robot AI models have their own schema and are counted separately.
- Stated by the manufacturer
- 3,875
- Not stated
- 13,681
- Possible pieces of information 33 fields × 532 robots
- 17,556
- Stated by the developer
- 670
- Not stated
- 646
- Possible pieces of information 14 fields × 94 robot AI models
- 1,316
The four states
A figure, a zero, a no and a gap are four different things, and each looks different on the page, so a zero is never shown as an empty cell and an empty cell is never read as a zero. In this catalogue the manufacturers state 28 measured zeros, and 36 times they answer no outright.
The robot pages use these same marks. Each figure also carries a source mark (A, B, C …) that opens its source and note.
- 14kg30.9lbA figure the manufacturer states.
- 0A reported value of zero, not missing data.
- NoExplicitly denied by the source.
- Not statedNo published source states this field.
Technical details
Source marksShowHide
3,852 of the 3,875 stated fields carry a source mark. 3,268 point to the manufacturer’s own product page, 584 to a secondary source – a datasheet, a manual or developer documentation from the manufacturer, or a press report where the manufacturer gives no figure. The remaining 23 are the answer no, which has no figure to source.
The steps in fullShowHide
- Fetched
- We fetch the manufacturer’s own page, datasheet or manual and keep the file as evidence, with its address and the date it was fetched.
- Recorded
- The figure goes into one of the schema’s 33 fields with a unit and a source. If the manufacturer does not state it, the field becomes not stated. It never becomes an estimate.
- Checked for completeness
- The validator reads the whole catalogue again and stops the build if a figure is missing a unit, a source or a date. It checks that each figure is complete and sourced, not that the manufacturer’s claim is true; we do not test the robots. If the manufacturer’s metric and imperial figures do not match, we never correct them quietly: either the build stops, or the disagreement stands on the page beside the figure.
- Built
- The pages are generated afresh from the same files, and the build runs the validator first every time. That is why the site carries 0 figures without a source.
Where the pictures come fromShowHide
Pictures are the manufacturers’ and developers’ own photos, and press photos. For an AI model, a still from a video of the model at work may stand in. Where a background has been removed, the robot itself is unchanged.
No picture of a robot or a person is AI-generated. The one exception is a world or simulation model’s own output, which may appear on that model’s own page and is marked “Generated by the model”. Models without their own picture show a real photo of robot training or use, marked “Illustrative image — not this model”.
Robot AI modelsShowHide
Some entries are AI models, not robots. A model’s page describes the software; the robots it runs on have their own pages. A model’s type is the one its developer calls it – vision-language-action (VLA), world action model, embodied reasoning or general-purpose model – and reads “Not stated” where the developer uses none of these words. A vision-language-action model takes camera images and an instruction and turns them into movements for a robot.
Figures on a model’s page come from its developer. A general-purpose model is included when its robot use is documented by the developer or by a named independent study; that study is a secondary source and shows only that the model has been used for robot tasks. The catalogue gives no benchmark results and no ranking of models.
Publisher
KeyResearch in Aarhus, Denmark, publishes the catalogue as a whole. It is not the source of any individual entry, and it is not a party to any sale.
For corrections or updated datasheets, contact the editorial team. Corrections are verified at the source with an updated retrieval date.
- Address
- KeyResearchIncuba, Åbogade 15, Aarhus
- Telephone
- +45 22231116
- jpk@keyresearch.dk
The maps on the manufacturers page are drawn from Natural Earth (naturalearthdata.com, public domain).
City positions come from Natural Earth and from GeoNames (geonames.org), whose data is used under the Creative Commons Attribution 4.0 licence (CC BY 4.0, creativecommons.org/licenses/by/4.0).