AI Atlas / Robotics systems
Robotics systems, feature by feature
Robots and vehicles in which a learned model perceives and acts in the physical world, and the models and platforms under them. Each row is one system; each column is one precisely defined feature. Every filled cell carries the page it was read from and the sentence that supports it. The software counterpart is Agentic systems.
First deployments by category
Each dot is one system, placed at the date of its first use by an outside customer or the public, or its first sale. Select a dot for the system.
- Date from the maker, a regulator or a paper
- Date from reporting, or unverified
The same as a table
Loading…
Feature matrix
Select a system's name for its sources, quotes, timeline and figures. Select a column heading to keep only the systems that have that feature. Scroll sideways for more columns.
- Yes meets the definition
- Partial with a stated limit
- No a page says so, or its description of the robot's body shows it
- Searched, not found the system's own pages were searched for this and nothing was found; it may still exist
- Unknown the sources could not be read
- No letter: read from the maker's, a regulator's or a paper's own text.
- From reporting, not the maker named news reporting, or a partner's statement
- Unverified, pending source confirmation a single thin source, or a claim that could not be traced
Reviewed on 6 and 8 October 2026; still a hand-built record. On 6 October every one of the 332 "partial" cells in the first 61 systems was re-read against its source page: 235 stand, each now naming the limit that applies, and 97 were changed (48 to "yes", 47 to "searched, not found", 2 to "no"). A fresh sample of 64 cells was then read against the pages: 1 was wrong and 1 arguable. The fifteen systems added after that (AgiBot, XPENG IRON, Sanctuary Phoenix, Amazon Sparrow and Astro, Symbotic, Geek+, Coco, Matic, DJI Agras, Monarch Tractor, the John Deere autonomous tractor, Mobileye Drive, Avride and TuSimple) were read cell by cell on 8 October. Of their 221 cells that carried a value, 21 were given a different one (about 1 in 10, which is why that review was done), 39 "not found" cells turned out to have an answer on pages the first pass had not read, and several first-demo and first-deployment dates were corrected. A fresh random sample of 64 of the 260 value-carrying cells in those fifteen, read by a separate checker, found none wrong and 5 arguable; 19 of the 64 are on pages that checker could not read (mostly SEC filings, which block unnamed tools) and were judged from the quote alone, so read "none wrong" as "none found". The five cells it called arguable were then settled on 10 October: two were lowered to "partial" (Matic's regulatory approval, Symbotic's navigation), two were moved to "searched, not found" (Avride's operator assist, Deere's on-site person) and one was kept (AgiBot's regulatory approval). Regulators' own recall, investigation or product-safety records were read for 21 of the 76 systems; for 38 they were not read, and there a list with no incident or recall is a gap in this record, not a clean one; 17 are research models or platforms with nothing a regulator would hold. Each system's panel says which. 208 of the 2,757 statements that cite a page rest on pages the site's automatic check cannot re-read, and are labelled.
Patterns
How far to trust these
Autonomy by year of first deployment
For the systems with a dated first deployment: how many have each autonomy feature today, grouped by when they were first deployed. The values are today's; this does not show what a system could do when it was first deployed.
How the categories differ
For each category with at least four systems: the features, other than the body itself, whose share differs most from the rest of the set.
Features that almost always come together
Each feature is present in at least five systems, at least three systems have neither, and at least four in five of the systems that have either have both.
Features that almost never appear together
The test for one feature excluding another: each is present in at least five systems, yet at most one system has both. Exclusions that come from the body (an aircraft does not walk) are labelled and listed last.
One feature that comes with another
Systems with the first feature have the second, with at most one exception, while several systems have the second alone and several have neither.
Recurring combinations
How common each feature is
Bar: share of systems with a known value that have the feature; the lighter part is partial. The figures give the count, then how the rest divide between confirmed absent and a search that found nothing, then how many of the cells are not verified.
Changes
The set is rechecked on a rolling basis. Each entry says what was added or changed and why.