Physical AI is becoming a modular stack, not a humanoid race
The visible market is becoming less defined by robot form. Growth is shifting toward manipulation, vision-language-action models, simulation, and shared platforms that can transfer across physical systems.
The platform layer is expanding as humanoid share contracts
Within the broad GitHub cohort, humanoid projects fell from 17.7% to 9.3%. VLA, world-model, simulation, and development-platform projects rose from 79.3% to 82.4%. Modular projects not tied to humanoids reached 79.6% of the cohort.
Robot form is giving way to a shared stack
Shares use only the broad robotics, embodied AI, and physical AI cohort, excluding category-specific queries from the denominator.
-8.4 pp
+3.1 pp
+5.5 pp
comparable quarters
GitHub: manipulation and VLA grow despite age bias
Categories are multi-label; one project may appear in multiple rows.
Q2 projects had less time to reach the three-star threshold. The absolute humanoid decline is therefore not proof of market contraction, while the relative redistribution within the broad cohort remains informative.
Hugging Face shows an explosion in VLA supply
Artifacts, unique publishers, and models with non-zero downloads are reported separately so checkpoint variants are not mistaken for independent projects.
+591.8%
367%
20,998 downloads
683 unique
GitHub supply remains persistent
Unique projects after deduplicating category queries.
263 projects in Q1 and 229 in Q2. June cannot be compared with older months by stars alone because newer repositories need time to cross the inclusion threshold.
The research filter supports the platform direction
The current HF Daily Papers snapshot is used only as an independent qualitative check.
current snapshot
papers
papers
paper
What this means for the market
Value is moving into transferable layers
VLA, simulation, data, and platforms can serve multiple robot bodies, reducing dependence on a single form factor.
Manipulation is closer to applied demand
Manipulator growth points to warehouse, manufacturing, and laboratory tasks where value comes from completed work rather than human likeness.
VLA is becoming a shared interface
The sharp rise in Hugging Face publishers shows growth not only in checkpoint variants but also in independent teams working on the approach.
Signals tested by this research
A direct link between AILANTA's early observations and the study's measured findings.
GitHub combines broad and category searches from January through June 2026, capped at two top-by-stars pages per query and month. Hugging Face uses models created in the same period and classifies them from metadata and README text. Product Hunt is not used.
Categories overlap and do not sum to 100%. GitHub measures visible open-source supply and has an age bias from stars. Hugging Face downloads and likes are current snapshots rather than historical values. HF Daily Papers is a current community filter, not a six-month census.