EN

X-VLA Team

X-VLA 0.9B

X-VLA 0.9B

Checkpoint

Soft-prompted flow-matching VLA trained across robot embodiments.

Specifications

Architecture components: Flow-matching action expert
Architecture components: Embodiment soft prompts
Backbone: Florence-2 vision-language backbone with soft prompts and flow matching
Frameworks: LeRobot / PyTorch
Input modalities: Language instruction
Input modalities: RGB camera images
Input modalities: Robot state / proprioception
Languages: English
Model type: Vision-language-action model
Output action space: Continuous action chunks up to 20 dimensions
Parameter count: 0.9B parameters
Pipeline task: Robotics
Training data scale: 290K episodes

Lineage

Unknown

Architecture & I/O

trajectory · continuous

Unknown

Continuous action chunks up to 20 dimensions

Source · lerobot

Default checkpoint

lerobot/xvla-libero

openApache License 2.0

Other releases

6

published checkpoints

lerobot/xvla-base lerobot/xvla-agibot-world lerobot/xvla-folding lerobot/xvla-google-robot lerobot/xvla-libero lerobot/xvla-widowx

Supported robots

No verified robot compatibility recorded.

Benchmark results

LIBERO · LIBERO average

Success rate: 93 percent

official · reported · Protocol LeRobot 400-episode protocol

Success rate reported by the official LeRobot X-VLA documentation.

Source · Hugging Face LeRobot

Training data

finetune

HuggingFaceVLA LIBERO

open

Source · Hugging Face LeRobot

pretrain

open

Source · Hugging Face LeRobot

finetune

open

Source · Hugging Face LeRobot

finetune

RT-1 Robot Action

open

Source · Hugging Face LeRobot

pretrain

open

Source · Hugging Face LeRobot

pretrain

open

Source · Hugging Face LeRobot

finetune

open

Source · Hugging Face LeRobot

Resources

documentation · LeRobot X-VLA documentation paper · X-VLA paper repository · X-VLA repository hub · lerobot/xvla-libero model card weights · lerobot/xvla-libero model card

Submit info