WorldVLA

Alibaba, China 路 June 2025

Alibaba DAMO Academy, Hupan Lab and Zhejiang University

World model 路 Open weights, Apache-2.0 路 Also written WorldVLA-ActionModel

Parameters
not published
No count in the paper.
GPU memory
not published
weights at bf16, computed
Inference latency
not published
per action step
Weights
Apache-2.0
Apache-2.0 according to the model card.

What it is

WorldVLA is an autoregressive action world model from Alibaba DAMO Academy that puts a VLA and a world model into one Chameleon-based transformer. Images, text and actions share one token vocabulary, so the same weights predict actions from images and an instruction and predict the next image from an image and an action. On LIBERO it reaches 81.8 percent average success at 512 by 512 pixels and 79.1 percent at 256, without robot pretraining. An attention mask that hides earlier generated actions while producing a chunk improves grasping success by 4 to 23 percent, according to the paper. The project continued as RynnVLA-002, and the GitHub repository now redirects there.

Architecture

Backbone
Chameleon (facebook/chameleon-7b according to the model card), a unified image and text token model. Image tokenizer with compression 16 and 8,192 codes, actions as 7 tokens with 256 bins per dimension, shared vocabulary of 65,536 tokens.
Action head
Discrete action tokens generated autoregressively, with an attention mask that hides previously generated actions within a chunk. The same weights predict future image tokens from an image and an action.
Parameters
No count in the paper. The model card names facebook/chameleon-7b as the base model. Checkpoints are released per LIBERO suite at 256 and 512 pixels.
Pretraining data
The LIBERO results are reported without robot pretraining, starting from Chameleon.
Embodiments
Franka Emika Panda (LIBERO simulation)

What hardware it needs

This is the section most people came for, so it is worth being precise about which numbers are measured and which are arithmetic.

Not published.

The authors publish no memory or latency figure for this model. Everything above is computed from the parameter count.

Reported results

Grouped by what the evaluation actually asked. Values from different suites are not comparable with each other, so each one keeps its suite and split.

Simulation

Reproducible benchmark suites. Everyone runs the same episodes, so the numbers can be compared inside a suite.

  • LIBERO average over Spatial, Object, Goal and Long, 512 by 512 pixels, without pretraining
    81.8%
    success rateAuthors' own evaluation. Per suite: Spatial 87.6, Object 96.2, Goal 83.4, Long 60.0.source
  • LIBERO average over Spatial, Object, Goal and Long, 256 by 256 pixels, without pretraining
    79.1%
    success rateAuthors' own evaluation. Per suite: Spatial 85.6, Object 89.0, Goal 82.6, Long 59.0.source
  • LIBERO Long suite, 512 by 512 pixels
    60%
    success rateAuthors' own evaluation. The hardest LIBERO suite and the most informative of the four.source

Real world

No results in this category are published for this model.

Fine tuned tasks

No results in this category are published for this model.

On real hardware

No real robot results are reported.

Fine tuning it yourself

Checkpoints are published per LIBERO suite at 256 and 512 pixels. The code moved to the RynnVLA-002 repository.

Where it helps, where it does not

Strengths

  • One set of weights acts as policy and as world model, and the paper reports that each part improves the other.
  • 81.8 percent average on LIBERO without robot pretraining.
  • Open checkpoints on Hugging Face.

Limits

  • Simulation only: no real robot evaluation.
  • LIBERO-Long reaches only 60.0 percent, against 94.4 percent for its successor RynnVLA-002.
  • Checkpoints are per-suite specialists, not one generalist model.
  • Licence question: the model card says Apache-2.0, while the Chameleon base model has a research licence.

Sources

Everything on this page was taken from these documents. Where they disagree with what you read here, they win.

Entry last checked 2026-11-12.

Try a policy on a real arm

A real SO-100 runs on /live and can be driven from the browser when it is online. Five of the models in this arena can be fine tuned on a dataset you record with your own arm.