VLA GalleryAll modelsGallery

VLA Gallery / All models / Autoregressive & latent-action VLA

WorldVLA

Alibaba DAMO Academy·Hupan Lab·Zhejiang University · 2025-06

VLAAutoregressive & latent-action VLANot discloseddiscrete tokens

Unifies action generation and next-image prediction (world model) in one autoregressive model.

View architecture diagram in the galleryPaper ↗

Paper

WorldVLA: Towards Autoregressive Action World ModelarXiv 2506.21539

Overview

Category
Autoregressive & latent-action VLA
Type
VLA
Released
2025-06
Organization
Alibaba DAMO Academy·Hupan Lab·Zhejiang University
Open weights
Yes
Parameters
Not disclosed (Chameleon based)

Architecture

Backbone LLM
Chameleon 7B
Vision encoder
VQ-GAN image tokens
Action head
256-bin tokens + image generation (discrete tokens)
Chunk
10 · 5
Control rate
Not disclosed
Architecture (original summary)
Chameleon initialization (README references the Lumina-mGPT-7B-768 tokenizer); VQ-GAN image tokens (codebook 8192) + 7 action tokens of 256 bins, chunk K=10 (LIBERO-Long)/5, action attention mask

Training

Pretraining accelerators
Not disclosed
Pretraining duration
Not disclosed
Pretraining steps & batch
Not disclosed
GPU/TPU-hours
Not disclosed
Fine-tuning
Not disclosed
Data type
Simulation·synthetic
Data scale
LIBERO demos (90/10 split after excluding failures)
Main datasets
LIBERO
Tasks & evaluation
LIBERO average 81.8% (512², OpenVLA 76.5%)
Robot
LIBERO simulation

Inference

Inference device & speed
About 3.67 FPS (1 frame, chunked)–1.22 FPS (4 frames, no chunking); GPU not stated
Inference memory
Not disclosed
Weights bf16 (calc.)
Not disclosed
Where it runs
Not disclosed

Reinforcement learning & base

Reinforcement learning
Not used · Supervised video·action generation
Base control
Fixed · No base control (LIBERO arm sim)

Notes

Key features
Unifies action generation and next-image prediction (world model) in one autoregressive model. Uses an attention mask that hides previous actions during chunk generation.
Notes
The README tokenizer path suggests Chameleon-7B, but the parameter count is not stated. The repo was upgraded to RynnVLA-002 in 2025-11.
Verification
Confirmed

Other models in this category