LeWorldModel · SADL shared-BatchNorm

This LeWorldModel research variant adds SADL to shared batch normalization training. It predicts future latent representations and uses CEM for planning. The one-epoch release includes joint and separate training on PushT, Cube, Reacher, and TwoRoom.

View full model description

Visual world modeling and goal-conditioned control in simulation.

Papers & reports (1)

SADL: On the Demand-Supply Geometry of Representation Collapse

Paper available

arXiv: 2611.21596DOI: 10.48550/arxiv.2611.21596
Authors & affiliations (10)

Author order: Original order

  1. Junhan Sun

    State Key Laboratory of CAD&CG, Zhejiang University

  2. Yigui Feng

    RoboParty Lab

  3. Mingzhe Liu

    RoboParty Lab

  4. Haoyan Xu

    RoboParty Lab

  5. Lei Ye

    RoboParty Lab

  6. Ziqi Han

    RoboParty Lab

  7. Hanru Bai

    Alibaba Group

  8. Yanzhao Zhang

    Alibaba Group

  9. Hao ZhaoCorresponding author

    RoboParty Lab · Institute for AI Industry Research (AIR), Tsinghua University

First published
2026-11-16
Latest paper revision
v1 · 2026-11-16
Venue
arXiv

Covered model versions

Paper ↗PDF ↗

How to access and use

Public code resources
Public access
Model files
Public access
Training data
Public access

Code and training / fine-tuning

Access
Public access

Verified on · 2026-11-22

Code license
MIT

Model files

Model files
E1 · Four tasks · seed3072AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

E1 · Four tasks · seed3073AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

E1 · Four tasks · seed3074AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

E1 · Cube · seed3072AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

E1 · PushT · seed3072AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

E1 · Reacher · seed3072AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

E1 · TwoRoom · seed3072AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

E1 · Cube · seed3073AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

E1 · PushT · seed3073AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

E1 · Reacher · seed3073AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

E1 · TwoRoom · seed3073AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

E1 · Cube · seed3074AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

E1 · PushT · seed3074AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

E1 · Reacher · seed3074AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

E1 · TwoRoom · seed3074AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

The author explicitly licenses the released weights under MIT. Keep the copyright and license notices when redistributing the weights. Datasets, simulator assets and runtime dependencies have their own terms.

Release information

SourceVerified on · 2026-11-22

Checks before deployment

The released checkpoints are evaluated on four simulation benchmarks. Hardware validation for these weights has not been reported in the release materials. Available weights or code do not establish plug-and-play robot compatibility. Check the required configuration, data and interfaces in the source documentation.

Deployment and usage resources

Architecture and inputs / outputs

Model type
World simulation
Architecture
Transformer
Input
Goal image / RGB images
Released
November 19, 2026
  • Input modalities · Main category · Architecture: github.com
  • Input modalities · Main category · Architecture: huggingface.co

Scope and training

Output

Representations

Task scope

Multiple tasks

Evaluation results

No verified independent benchmark results are recorded for this version or the selected weights.

Full parameters and verification notes (model card)

The ViT-Tiny/14 encoder starts from scratch. Observations are 224 × 224 RGB, with 192-dimensional latents. The joint training batch has 256 windows of 8 frames, 64 per task, and shares running BatchNorm statistics. CEM uses 300 samples per iteration for 30 iterations, the top 30 candidates, and initial standard deviation 1.0. Goals are 25 steps ahead. The action budget is 50, the planning horizon is 5 action blocks, and replanning occurs every 5 actions. Evaluation uses FP32, solver batch size 1, and an RTX 4090. Each training seed (3072–3074) is evaluated with seeds 0, 1, 42 and 100 episodes per evaluation seed per task. Mean ± standard deviation over training seeds: joint 66.33 ± 9.19%, separate 78.58 ± 1.89%. Joint / separate per task: PushT 64.67 ± 18.48 / 83.11 ± 1.58%, Cube 59.56 ± 1.26 / 58.89 ± 1.39%, Reacher 46.67 ± 26.12 / 81.00 ± 4.26%, TwoRoom 94.44 ± 0.96 / 91.33 ± 1.45%. Other paper ablations use distinct training runs.

Submit a source or correction