ModelFeatured
Octo
Generalist robot policy Transformer with multi-modal inputs and multi-embodiment finetuning.
93MFully openData: Open X-EmbodimentUpdated 2024-07-31
ModelFeaturedFully open
SmolVLA
Hugging Face / LeRobot lightweight VLA (~0.5B) for fast prototyping and low-GPU finetune loops.
Why we recommend it
SmolVLA aims at a smaller footprint for learning the VLA loop on limited GPUs. Prefer it when your goal is a closed finetune loop, not the largest leaderboard model. Still verify licenses and expect domain gap on real tasks. Beginner-to-intermediate.
Overview
SmolVLA refers to small VLA efforts in the LeRobot / Hugging Face ecosystem designed for accessible finetuning. The point is not SOTA leaderboard scores — it is fitting vision-language-action training on modest GPUs.
Follow LeRobot docs for dataset format, training entrypoints, and eval. Combine with OXE subsets or your own teleop data.
Learning Path (VLA): first finetune target before full OpenVLA / π0-scale models.
Who it is for
Researchers, students, and builders comparing open embodied stacks.
Key highlights
- Small VLA efforts in the LeRobot / HF ecosystem for accessible finetuning experiments.
- Type: model; categories: model
- Marked recommended on the hub
When to use
- You are mapping the open embodied AI landscape
- You need a starting URL and judgment note before deep-diving
When not to use
- You need production SLAs or certified industrial support only
- License or hardware constraints are incompatible with your deployment
Getting started
- 1Open the official URL and read README install pins.
- 2Skim license and citation requirements.
- 3Run the smallest official example before scaling experiments.
- 4Cross-link related hub resources (sim / data / policy) for a full stack.
Install / setup
via LeRobot ecosystem
pip install lerobot
How it compares
Caveats & pitfalls
- Hub cards are curated summaries — always verify upstream docs.
- Stars and dates drift; treat extra.stars as approximate.
Content reviewed 2026-07-28