WORLD-MODEL-AUGMENTED VLA EXECUTION

WAVE-0

Bridging VLA Policies and WAMs
with Lightweight Latent Future Prediction

Haomin Wang1 Anqiong Ji1,3 Ke Wang2,4 Zhongshan Liu3 Ru Ying3
Ran Zheng3 Zhipeng Zhang1†
1AutoLab, School of Artificial Intelligence, Shanghai Jiao Tong University 2Awaken Robotics 3Baidu AI Cloud 4Zhejiang University

†Corresponding Author

arXiv Code Models
RoboTwin 2.091.47%
LIBERO98.7%
LIBERO-Plus88.0%
LIBERO-Pro61.4%
Real world73.3%
Latent world model257M

01 THE IDEA

From understanding
the world to
anticipating it.

How can we bring WAM-style future prediction into pretrained VLA policies?

WAVE-0 bridges the semantic understanding of pretrained VLAs with future prediction in WAMs through a lightweight, 257M latent world model. Built on π0.5, it translates VLM-generated latent action plans into predicted future visual representations that condition the action expert alongside the original vision–language context, during both training and inference.

Our two-stage training first learns action-dependent latent dynamics directly from ground-truth robot actions, with counterfactual ranking encouraging the world model to distinguish the consequences of different actions. We then integrate the frozen world model into π0.5 and align VLM-generated plans with the learned action representations—unifying latent planning, future prediction, and action generation without pixel reconstruction or additional large-scale video pretraining.

Explore the architecture
Action-grounded dynamics learningLightweight latent future predictionFuture-conditioned action generation

02 THE METHOD

Plan. Predict. Act.

Latent action plans become predicted futures
that explicitly condition action generation.

  1. 01 / PLAN

    Latent action planning

    Given the current observation and instruction, the VLM uses learnable action queries to produce latent plans aligned with action representations learned in Stage 1.

  2. 02 / PREDICT

    Future world state prediction

    The frozen world model combines these plans with current visual representations and robot state to predict future representations in a fixed visual embedding space.

  3. 03 / ACT

    Future-conditioned action generation

    A resampler compresses predicted futures into tokens that condition the action expert alongside the original VLM context. Latent-action query tokens are excluded from the expert’s context.

Overview of WAVE-0. Stage 1: The latent world model learns action-conditioned dynamics in a fixed visual embedding space, trained with future-representation prediction, latent residual supervision, and counterfactual action ranking. Stage 2: The VLM generates latent action plans aligned with the representations from the frozen action encoder. The frozen world model predicts the visual consequences of these plans, and a resampler compresses them into future tokens that condition the action expert alongside the original VLM context.
STAGE 1

Learn how actions change the world.

We jointly train the action encoder, state encoder, and latent world model to predict future visual representations from current observations and ground-truth robot actions. Future prediction and latent residual supervision teach the model action-dependent dynamics, while counterfactual ranking encourages it to distinguish demonstrated actions from mismatched ones when predicting the observed future.

STAGE 2

Let predicted futures guide actions.

We integrate the frozen world model into π0.5 and train the VLM to generate latent action plans aligned with representations from the frozen action encoder. The world model predicts the visual consequences of these plans, which a lightweight resampler compresses into future tokens. These tokens condition the action expert alongside the original VLM context during both training and inference.

03 THE RESULTS

Small world model.
Meaningful gains.

Evaluated in simulation, under distribution
shifts, and on physical robots.

FOUR TASK SUITES98.7%

Stronger across
sequential tasks.

The largest gain over π0.5 is on LIBERO-Long: 92.4% to 97.6%, where earlier actions shape the success of later steps.

Spatial · Object · Goal · Long
Success rate (%) on LIBERO
MethodSpatialObjectGoalLongAvg.
OpenVLA84.788.479.253.776.5
GR00T N194.497.693.090.693.9
π096.898.895.885.294.1
Motus96.899.896.697.697.7
LingBot-VA98.599.697.298.598.5
Fast-WAM98.2100.097.095.297.6
ImageWAM97.299.298.898.498.4
VLA-JEPA96.299.697.295.897.2
JEPA-WAM95.699.497.294.696.7
LaWAM99.499.698.497.098.6
π0.598.898.298.092.496.9
WAVE-0 Ours98.6100.098.697.698.7

04 IN THE REAL WORLD

From latent futures
to physical actions.

Three tasks. Dual-arm AgileX Piper.
20 evaluation trials per task.

Book Organization

85.0% vs. 40.0% · π0.5

USB Storage

65.0% vs. 45.0% · π0.5

Keyboard Pushback

70.0% vs. 30.0% · π0.5

Real-world performance

73.3% average +35.0 pp over π0.5

Real-world success rates (%)
MethodBook OrganizationUSB StorageKeyboard PushbackAvg.
ImageWAM50.025.045.040.0
LaWAM40.035.055.043.3
π0.540.045.030.038.3
WAVE-0 Ours85.065.070.073.3
Δ Improvement (pp)+45.0+20.0+40.0+35.0

05 BUILD WITH WAVE-0

A lightweight idea.
Ready to explore.

Training recipes and model checkpoints
for WAVE-0.

GITHUB

Explore the code

Implementation, two-stage training,
and evaluation scripts.

HUGGING FACE

Model collection

WAVE-0 checkpoints for
LIBERO and RoboTwin 2.0.

QUICK START

Start experimenting

Prepare your environment
and run the training recipes.

git clone https://github.com/AutoLab-SAI-SJTU/WAVE-0.git

Citation

@misc{wang2026wave0bridgingvlapolicies,
      title={WAVE-0: Bridging VLA Policies and WAMs with Lightweight Latent Future Prediction},
      author={Haomin Wang and Anqiong Ji and Ke Wang and Zhongshan Liu and Ru Ying and Ran Zheng and Zhipeng Zhang},
      year={2026},
      eprint={2610.18568},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2610.18568},
}