ModelFeatured
π0 (pi0)
Physical Intelligence generalist robot foundation model with flow-matching action generation.
ModelPaper only
RT-2
Google DeepMind VLA that transfers internet vision-language knowledge into robot control.
Overview
RT-2 is Google DeepMind’s seminal Vision-Language-Action work that casts robot actions as text-like tokens produced by a VLM backbone. It established the modern VLA narrative: web-scale semantic knowledge transferred into robot control.
Public reproducibility is limited compared with OpenVLA; treat RT-2 primarily as a conceptual and paper reference unless you have access to internal stacks.
Who it is for
Readers studying VLA history and design; survey authors.
Key highlights
- Defined the VLA paradigm for many follow-ups
- Web-scale semantics → control story
- Influential evaluation narrative
When to use
- Literature review, design inspiration, citation baseline
When not to use
- You need fully open weights for local finetune today
Getting started
- 1Read the paper carefully
- 2Compare claims with OpenVLA/Octo open stacks
- 3Reproduce ideas, not proprietary weights
Papers & reading
How it compares
ModelFeatured
NVIDIA GR00T N1
NVIDIA foundation model series for humanoid and general robots combining sim and real data.
Weights openData: NVIDIA 多源人形与操作数据Updated 2026-03-01
Caveats & pitfalls
- Not a drop-in open checkpoint for most labs.
Content reviewed 2026-07-23