RoboAtlas
EN
← VLA-JEPA

VLA-JEPA Authors

VLA-JEPA

VLA-JEPA

Checkpoint

VLA with Qwen3-VL, a flow-matching action head and a training-time V-JEPA2 world model.

Specifications

Architecture components
Flow-matching action expert
Source · lerobot
Architecture components
World-action model
Source · lerobot
Backbone
Qwen3-VL-2B + DiT action head + V-JEPA2 world model
Source · lerobot
Frameworks
LeRobot / PyTorch
Source · lerobot
Input modalities
Language instruction
Source · lerobot
Input modalities
RGB camera images
Source · lerobot
Input modalities
Robot state / proprioception
Source · lerobot
Languages
English
Source · lerobot
Model type
Vision-language-action model
Source · lerobot
Model type
World-action model
Source · lerobot
Output action space
Continuous 7-DoF action chunks and latent future predictions during training
Source · lerobot
Parameter count
3.0815B parameters
Source · lerobot
Pipeline task
Robotics
Source · lerobot

Lineage

uses backbone

Qwen3-VL-2B
Source · Hugging Face LeRobot

uses world encoder

V-JEPA 2
Source · Hugging Face LeRobot

Architecture & I/O

trajectory · continuous

7 DoF

Continuous 7-DoF action chunks and latent future predictions during training

Source · lerobot

Default checkpoint

lerobot/VLA-JEPA-Pretrain

openApache License 2.0
Checkpoint

Other releases

3

published checkpoints

lerobot/VLA-JEPA-Pretrainlerobot/VLA-JEPA-LIBEROlerobot/VLA-JEPA-SimplerEnv

Supported robots

No verified robot compatibility recorded.

Benchmark results

No published benchmark results.

Deployment

Training data

Resources

documentation · LeRobot VLA-JEPA documentationpaper · VLA-JEPA paperweights · lerobot/VLA-JEPA-Pretrain model card
Submit info