W²-VLA
- AI model
- Humanoids and robot learning
A vision-language-action model that predicts future wrist latents and outputs robot actions for fine-grained manipulation, trained and evaluated on LIBERO, RoboTwin and real-world LeRobot data.
Source on GitHub: github.com/yyyyu120/W2-VLA (opens in a new tab). Facts read at commit e63f90c (opens in a new tab).
Licenses
- Code
- MIT (opens in a new tab)
Read by Robotheca from LICENSE (opens in a new tab) (MIT) at commit e63f90c (opens in a new tab), on 16 Oct 2026.
- Commercial use
- Yes
What it does
- Combines a Qwen3-VL world branch with a frozen V-JEPA2.1 wrist encoder, a latent predictor for future wrist frames, and a DiT action head.
- Uses offline chain-of-thought labels during training; inference uses modeling-token states without generating text.
- Provides training scripts for LIBERO, RoboTwin and four real-world tasks, plus policy-server evaluation for LIBERO and RoboTwin.
- Released checkpoints and datasets are hosted on Hugging Face; the training data release contains 58 LeRobot datasets.
Written by Robotheca from the project's README (opens in a new tab) at commit e63f90c.
Getting started
- Difficulty
- Advanced
- How to start
- The project's README (opens in a new tab)
Maintenance
- Status
- Active: a commit in the last 90 days
- Newest commit
- 2 Oct 2026 on main
- Latest release
- W2-VLA v1 (5 Aug 2026)
Checked on GitHub on 16 Oct 2026.
Links
Robots that can run it
No robot on Robotheca is linked to it yet. Links come from a robot's robot.yaml, from its dependency files read as text, or from Robotheca's editors.
Robotheca lists AI models as a reference: nothing is hosted or run here, and every fact links to where it was read.