RynnVLA-002

Alibaba, China · November 2025

Alibaba DAMO Academy

World modelOpen weights, Apache-2.0
Parameters
not published
No total parameter count is stated in the paper or the repository.
GPU memory
not published
weights at bf16, computed
Inference latency
not published
per action step
Weights
Apache-2.0
downloadable checkpoint

What it is

RynnVLA-002 is Alibaba DAMO Academy's November 2025 follow-up to RynnVLA-001, and it unifies a VLA and a world model in one Chameleon-based architecture with four tokenizers (image, text, state, action) over a shared 65536-token vocabulary. The world model predicts future image states from actions and observations, the VLA predicts actions from observations, and the paper's contribution is showing the two improve each other rather than merely coexisting. It reports 97.4 percent average on LIBERO with continuous actions and no robot pretraining, and on a real LeRobot SO100 the world model raises overall success by 50 percent relative to the VLA alone. A practical caveat: the released checkpoints are per-LIBERO-suite specialists rather than a single generalist model, and no parameter count, VRAM or latency figure is published anywhere.

Architecture

Backbone
Chameleon-based unified image understanding and generation architecture, using Lumina-mGPT-7B-768 components for tokenization
Action head
Hybrid: discrete action tokens plus a dedicated continuous Action Transformer head. Action attention masking isolates each action's dependence on visual and language input rather than on previously generated actions, which the paper introduces to stop error accumulation during sequential action generation.
Parameters
No total parameter count is stated in the paper or the repository. The system builds on Chameleon tokenizer and base-model components and uses the Lumina-mGPT-7B-768 architecture for tokenization, with four tokenizers (image, text, state, action) sharing a 65536-token vocabulary, but no aggregate figure is published.
Pretraining data
The headline LIBERO results are reported without robot pretraining, which the paper presents as the point. Released VLA checkpoints are at 256x256 resolution and world-model checkpoints at 512x512.
Embodiments
LeRobot SO100 arm (real-world experiments), LIBERO simulation suites

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.

Nothing published. Neither the paper nor the repository states GPU memory or compute requirements.

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, continuous actions, without pretraining
    97.4%
    success rateAuthors' own evaluation. Per-suite: Spatial 99.0, Object 99.8, Goal 96.4, Long 94.4 percent. Achieved without robot pretraining, which is the paper's headline framing.source
  • LIBERO Long suite, continuous actions
    94.4%
    success rateAuthors' own evaluation. LIBERO-Long is the discriminative suite, so this is more informative than the average.source
  • LIBERO average over the four suites, discrete actions
    93.3%
    success rateAuthors' own evaluation. Per-suite: Spatial 94.2, Object 96.8, Goal 94.6, Long 87.6 percent. The 4.1 point gap to the continuous-action variant quantifies what the Action Transformer head buys.source

Real world

Rollouts on physical hardware. The setups differ, so read these as evidence, not as a ranking.

  • Real robot (LeRobot SO100) overall, world model added versus VLA alone
    50%
    relative improvement in success rateAuthors' own evaluation, reported as a relative gain in the abstract. The absolute base success rate for the comparison is not stated there.source

Fine tuned tasks

No results in this category are published for this model.

On real hardware

Real-world LeRobot SO100 experiments report that adding the world model raises the overall success rate by 50 percent relative to the VLA alone. Per-task figures include 90 percent on multi-target block placement and 80 percent on multi-target strawberry placement, with 50 to 80 percent under distractors, reported as outperforming GR00T and pi-0 in cluttered scenes.

Fine tuning it yourself

Separate checkpoints are released per LIBERO suite (Spatial, Object, Goal, Long) for both the VLA at 256x256 and the world model at 512x512, so the released artefacts are task-specialised rather than a single generalist checkpoint. Training uses the LeRobot framework. Apache-2.0 across all assets and code unless otherwise specified.

Where it helps, where it does not

Strengths

  • 97.4 percent LIBERO average with continuous actions and without any robot pretraining, which is the strongest claim of the set because it removes the pretraining-corpus confound
  • Bidirectional design is empirically justified rather than asserted: the world model improves the VLA's real-world success by 50 percent, and the VLA improves the world model's visual consistency across viewpoints
  • Action attention masking directly addresses error accumulation in autoregressive action generation, a concrete and reusable fix
  • Apache-2.0 with released checkpoints for both the VLA and the world model
  • Real-robot results on the low-cost LeRobot SO100, so reproducible outside a corporate lab

Limits

  • No published total parameter count, VRAM, latency, control frequency or action chunk size
  • Released checkpoints are per-LIBERO-suite specialists at 256x256 (VLA) and 512x512 (world model), not one generalist model, so the 97.4 percent average is an average over four separately trained models
  • LIBERO at 97.4 percent is close to saturation and says little about hard real-world behaviour
  • Real-world evidence is limited to a single-arm SO100 with block and strawberry placement tasks
  • The 50 percent real-world improvement is reported as a relative gain, so the absolute base rate matters and is not given in the abstract

Sources

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

Entry last checked 2026-08-11.

Try a policy on a real arm

A physical SO-100 is online and free to drive in the browser, and five of the models in this arena can be fine tuned on a dataset you record with your own.