H Humanoid World

Knowledge

Robot Foundation Models — the 2026 landscape

Since 2024 a race has been on for the “brain” of humanoid robots. This overview maps the most important robot foundation models with their current status (mid-2026). Vendor performance superlatives are often self-reported and not yet independently verified — treated cautiously here.

At a glance

ProviderLatest modelStatusTypeLicense
Google DeepMindGemini Robotics 1.5 / ER 1.62025–2026VLA + reasoning VLMproprietary
NVIDIAIsaac GR00T N1.7 (N2 preview)2026VLA / world-actionopen
NVIDIACosmos (Predict/Transfer/Reason)2025–2026World modelopen
Physical Intelligenceπ0.72026VLApartly open (π0)
Figure AIHelixsince 2025VLA (System 1/2)proprietary
TeslaOptimus control net (unnamed)end-to-endproprietary
Skild AISkild Brain2026“omni-bodied” foundationproprietary
1X TechnologiesRedwood + World Modelsince 2025VLA + world modelproprietary
Hugging FaceSmolVLA / LeRobotsince 2025VLA / ecosystemopen
ByteDanceGR-32025VLA (Mixture-of-Transformers)research
AgiBotGO-1 (ViLLA)2025embodied foundationdataset open
World LabsMarble2026large world modelproprietary

The big picture

Proprietary & integrated (Google, Figure, Tesla, 1X): Companies that build their own robots or own a platform mostly keep the model closed and tune it tightly to their hardware. Gemini Robotics adds a separate reasoning model (ER) that “thinks before acting”.

Open & cross-platform (NVIDIA, Hugging Face): NVIDIA positions GR00T and Cosmos as open building blocks for the whole industry — flanked by its chip and simulation business. Hugging Face aims at democratisation with LeRobot/SmolVLA: small, open VLAs that run on consumer hardware.

Pure model startups (Physical Intelligence, Skild AI): They don't sell a robot body but the “brain” — a model that controls as many robot bodies as possible. Physical Intelligence's π0 is a widely noted, partly open VLA; Skild AI pursues an “omni-bodied” model.

The China bloc (AgiBot, ByteDance, Fourier, Unitree …): Chinese players combine aggressive hardware mass-production with their own embodied models and partly open datasets (AgiBot World). ByteDance's GR-3 and AgiBot's GO-1 are among the most visible.

The shared bottleneck

They all share one problem: too little robot data. Answers range from massive teleoperation and shared datasets (Open X-Embodiment) to synthetic data from world models (NVIDIA's “GR00T-Dreams”/Cosmos). This is where it will be decided whose model becomes reliable enough for everyday use first.

A more detailed, interlinked version (incl. individual model pages) is currently available in German: Foundation-Model-Landschaft.