VITRA
- AI model
- Humanoids and robot learning
VITRA is a pretraining approach for vision-language-action robot manipulation models, learned from large-scale real-life human hand activity videos treated as dexterous end-effector data.
Made by Microsoft, United States (checked on github.com/microsoft (opens in a new tab)).
Source on GitHub: github.com/microsoft/VITRA (opens in a new tab). Facts read at commit b355172 (opens in a new tab).
Licenses
- Code
- MIT (opens in a new tab)
Read by Robotheca from LICENSE.txt (opens in a new tab) (MIT) at commit b355172 (opens in a new tab), on 16 Oct 2026.
- Commercial use
- Yes
What it does
- Converts unannotated in-the-wild egocentric human hand videos into image, instruction and action data aligned with robotic training formats.
- Provides a human hand vision-language-action dataset of over one million episodes, named VITRA-1M.
- Offers a 3B-parameter base model, VITRA-VLA-3B, with a causal action transformer for zero-shot human hand action prediction from a single image.
- Supports few-shot fine-tuning on custom robot datasets, including an XHand dexterous hand example.
- Includes a teleoperation fine-tuning dataset released on Hugging Face.
- Includes scripts for real-world deployment and for pretraining from scratch on human data.
Written by Robotheca from the project's README (opens in a new tab) at commit b355172.
Getting started
- Difficulty
- Advanced
- How to start
- The project's README (opens in a new tab)
Maintenance
- Status
- Maintained: a commit in the last 12 months
- Newest commit
- 12 Jun 2026 on main
- Latest release
- None published on GitHub
Checked on GitHub on 16 Oct 2026.
Links
Robots on Robotheca that use 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.