INTACT · SADL shared-BatchNorm

This research variant applies SADL to the INTACT latent world model with shared batch normalization. The one-epoch releases cover joint and separate training on PushT, Cube, Reacher, and TwoRoom. The model predicts future representations and generates goal-conditioned actions with Direct or Guarded A planning.

View full model description

Simulation research on visual world models and goal-conditioned planning.

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

SourceVerified on · 2026-11-22

Code license
MIT

Model files

Model files
E1 · Four tasks · seed3072AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

E1 · Four tasks · seed3073AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

E1 · Four tasks · seed3074AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

E1 · Cube · seed3072AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

E1 · PushT · seed3072AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

E1 · Reacher · seed3072AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

E1 · TwoRoom · seed3072AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

E1 · Cube · seed3073AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

E1 · PushT · seed3073AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

E1 · Reacher · seed3073AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

E1 · TwoRoom · seed3073AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

E1 · Cube · seed3074AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

E1 · PushT · seed3074AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

E1 · Reacher · seed3074AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

E1 · TwoRoom · seed3074AccessWeights licenseCommercial use
ConditionalTerms

SourceVerified on · 2026-11-22

Conditions

MIT permits commercial use of the released weights. Keep the copyright and license notices in copies or substantial portions. Third-party components remain under their own terms.

Release information

SourceVerified on · 2026-11-22

Checks before deployment

The reported scores cover four simulation tasks. The release provides no hardware validation for these checkpoints. 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

Runtime and frameworks
PyTorch

Architecture and inputs / outputs

Model type
World simulation
Input
Goal image / RGB images
Released
November 19, 2026

Scope and training

Code license

MIT

Output

Actions / Representations

Task scope

Multiple tasks

Evaluation results

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

Publisher evaluation · Reported; not independently reproduced

SADL Shared BatchNorm Simulation · Joint training · Direct ↗

Macro success rate: 94.8 %

shared-batchnorm-e1

NVIDIA RTX 4090

FP32; solver batch size 1; last five primitive actions supplied. Direct uses zero candidates.

Joint training · Direct. Four simulation tasks, averaged across three training seeds (standard deviation 0.4 percentage points).

Publisher evaluation · Reported; not independently reproduced

SADL Shared BatchNorm Simulation · Joint training · Guarded A ↗

Macro success rate: 95.1 %

shared-batchnorm-e1

NVIDIA RTX 4090

FP32; solver batch size 1; last five primitive actions supplied. Guarded A: 128 samples × 3 iterations, top 16, initial SD 0.25.

Joint training · Guarded A. Four simulation tasks, averaged across three training seeds (standard deviation 0.3 percentage points).

Publisher evaluation · Reported; not independently reproduced

SADL Shared BatchNorm Simulation · Separate training · Guarded A ↗

Macro success rate: 95.6 %

shared-batchnorm-e1

NVIDIA RTX 4090

FP32; solver batch size 1; last five primitive actions supplied. Guarded A: 128 samples × 3 iterations, top 16, initial SD 0.25.

Separate training · Guarded A. Four simulation tasks, averaged across three training seeds (standard deviation 0.9 percentage points).

Full parameters and verification notes (model card)

Inputs are 224 × 224 RGB observations, a goal image, and the causal history of the previous five primitive actions. The ViT-Tiny encoder starts without pretrained weights and uses a 192-dimensional latent space. The evaluation uses three training seeds (3072–3074) and three evaluation seeds (0, 1, 42), with 100 episodes per evaluation seed and task on an RTX 4090. Goals are 25 steps ahead, with at most 50 actions, a planning horizon of 5 and replanning every 5 actions. Four-task macro success is 94.8 ± 0.4% (Direct) and 95.1 ± 0.3% (Guarded A). Separately trained models score 94.9 ± 0.5% and 95.6 ± 0.9%, respectively. The values are means ± standard deviations across training seeds. Inference uses FP32 with solver batch size 1. Direct generates the actor’s plan without candidate search. Guarded A evaluates 128 samples per iteration for three iterations, using the top 16 and an initial standard deviation of 0.25. These numbers describe the released Shared BatchNorm checkpoints; the paper’s other ablation protocols report separate results.

Submit a source or correction