OPEN MODELS/VISION-LANGUAGE-ACTION/X-VLA
RESEARCHVISION-LANGUAGE-ACTION

X-VLA

Soft-prompted VLA for diverse robot morphologies

WHAT DOES THIS ACTUALLY DO?

X-VLA solves a key problem in robotics: most models only work on one type of robot body. This model uses clever 'soft prompts' — small adapters — that teach it how to work with completely different robot shapes without retraining from scratch.

PARAMETERS
7B
RELEASED
2024-12
LICENSE
Apache 2.0
ARCHITECTURE
Soft-prompt...
CAN I ACTUALLY RUN THIS?
3
Research Only
OUT OF 10

This is a research model that requires significant compute. Realistically only accessible via cloud A100/H100 instances. Not suitable for local experimentation.

DOWNLOAD WEIGHTS
View on GitHub
15GB disk space required
OVERVIEW

A soft-prompted VLA framework using learnable embeddings to adapt to diverse robot morphologies and data types without full retraining.

TRAINED ON

Multi-embodiment datasets across diverse robot types

ARCHITECTURE
Soft-prompt adapter + VLA backbone
CAPABILITIES
Cross-morphology adaptation
Learnable soft prompts
Zero-shot transfer
Flexible input modalities
REAL ROBOTICS USE CASES
FLEET ROBOTICS

One model controlling arms, drones, and mobile bases

RESEARCH

Cross-embodiment generalization benchmarks

HARDWARE REQUIREMENTS
MIN VRAM
16GB
REC VRAM
40GB
MIN RAM
64GB
DISK SPACE
15GB
JETSON
Not Supported
CPU ONLY
GPU Required
TESTED GPUs
A100H100
DEPLOYMENT PATHS
Local PC
Requires 40GB+ VRAMNOT SUPPORTED
Cloud (A100)
SUPPORTED
Docker
SUPPORTED
ROS2
NOT SUPPORTED
Jetson
NOT SUPPORTED
Isaac Sim
NOT SUPPORTED
RELATED PAPERS
X-VLA: Learning a Universal Visual-Language-Action Model from Cross-Embodiment Datasets
RELATED ORGANIZATIONS
Hugging Face
ORGANIZATIONS
Hugging FaceResearch Community
DIFFICULTY
RESEARCH

Lab-grade infrastructure needed

COMPUTE1/10
SETUP2/10
BEGINNER-FRIENDLY2/10
EDGE DEPLOYMENT
Jetson Not Supported
CPU-only Not Available
Community →

Seeded and planned prompts — not live forum activity yet.