X-

X-VLA

0.9B (X-VLA-Pt)

A 0.9B generalist vision-language-action model built on flow matching and a soft-prompted Transformer, controlling robots across many embodiments.

✓ Active✓ Public access⚖ Open sourceVision-Language-Action modelRobotics foundation model

Parameters

0.9B

parameters

Release date

11 October 2025

Access:DownloadDeployment:💻 Local

Overview

X-VLA is a generalist Vision-Language-Action (VLA) model for robot control across diverse hardware platforms. Its architecture relies exclusively on a standard, soft-prompted Transformer with an action denoiser trained via flow matching. The model has 0.9B parameters and runs in BFloat16 precision.

Its key innovation is a separate set of learnable embeddings (soft prompts) for each distinct embodiment, letting the model exploit cross-embodiment features while staying scalable. It comprises a Florence-2 vision-language encoder, a SoftPromptedTransformer (a flow-matching action denoiser), and an Action Hub that manages action spaces, masking, and preprocessing.

X-VLA was trained on heterogeneous datasets (including Bridge Data) and evaluated in six simulations and on three real-world robots, achieving state-of-the-art results on the LIBERO, SimplerEnv, CALVIN, and RoboTwin 2.0 benchmarks. The weights are released under the Apache 2.0 license.

Classification

Vision-Language-Action modelRobotics foundation model

Access & deployment

Download

Local

Weights: Open source

Key parameters

🧩 Parameters: 0.9B

✓ Fine-tuning

📥 Input: image, text, robot state data

Technical specification

Parameters

0.9B

parameters

License

Apache 2.0

Hardware requirements

A single consumer/server-class GPU (0.9B model, BFloat16); PyTorch + Transformers.

Features:✓ Fine-tuning

Modalities

⬇ Input

imagetextrobot_state_data

⬆ Output

robot_actionsmotion_trajectories

Capabilities and applications

Native model capabilities

Vision-language-action grounding

The ability of a VLA model to ground visual perception and a language instruction into a concrete physical robot action. The model understands the scene and intent, then generates an executable action sequence, closing the loop from observation to motion.

Category: robotics

Cross-embodiment transfer

The ability of a single model to control robots with different morphologies (humanoids, dual-arm rigs, mobile platforms) without training a separate model per platform. Intelligence is decoupled from embodiment, so the same policy runs on hardware with different kinematics and dynamics.

Category: robotics

Action conditioning

Controlling model generation via action signals (camera, robot pose, commands, speech) rather than text prompts alone.

Category: multimodal

Instruction following

Precisely following instructions contained in the prompt: response format, length, style, constraints (e.g. 'reply in six words'). GPT-5.1 significantly improved this capability compared to GPT-5.

Category: language

Sample efficiency

The ability to reach strong performance using far fewer environment interactions or training examples.

Category: other