Embodied AI Benchmark Atlas

The Embodied AI Benchmark Atlas is a public reference site on how AI systems that control robots are tested. It covers 79 benchmarks and 70 world-model systems for readers who are new to the field.

Role
Research, product and design
With
VLGE
Timeline
2026
Links
Open the Atlas ↗
The Atlas home page
01 · Background

New embodied AI benchmarks appear every month, and most summaries do not show how well each claim is supported.

Embodied AI covers models that act in the physical world, such as robot arms that pick up objects. Labs measure these models with benchmarks, which are standard tests with published scores. Product managers, investors and students who are new to the field have no single place to check what each benchmark tests and whether its results hold on real robots.

I first mapped 32 benchmarks for VLGE in a research project. A table on a portfolio page is out of date within months, so I turned the research into a public site that stays current.

02 · An entry

Each entry shows the evidence behind every fact it states.

Every value carries a label: checked at the source, inferred, secondary or unknown, with the source number beside it. The site’s own assessment is kept separate and marked as opinion. Where sources disagree, both views are shown, and each entry records when it was last checked and when the next check is due.

The LIBERO entry: last checked, facts checked at the source, and the next check date
The LIBERO entry: last checked, facts checked at the source, and the next check date.
What a score on this benchmark does not show, stated at the top
What a score on this benchmark does not show, stated at the top.
A source card opened next to the fact it supports
A source card opened next to the fact it supports.
The site’s own assessment, labelled as opinion
The site’s own assessment, labelled as opinion.
A disputed claim, shown with the other view
A disputed claim, shown with the other view.
03 · Method

Only primary sources count as evidence, and every entry is checked again every 180 days.

Primary sources are the paper, the official site, the code or data, and the official leaderboard. News and other surveys are used only to find them. A weekly scan lists new papers and changed sources, and I review the list before the site changes. On the home page, the benchmarks appear as squares that regroup by where each test runs and whether it was checked against real robots.

The map of benchmarks by skill and by where each test runs
The map of benchmarks by skill and by where each test runs.
The home page squares regrouping as the page scrolls.
Next projectStability