LeWorldModel · LeWM

LeWorldModel is a JEPA world model trained end to end from images. It learns actions’ effects in a compact latent space and uses cross-entropy planning to reach visual goals in simulated control tasks.

Visual world modeling and goal-conditioned control in simulation.

Papers & reports (1)

LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels

Paper available

arXiv: 2603.19312DOI: 10.48550/arxiv.2603.19312
Authors & affiliations (5)

Author order: Original order

  1. Lucas Maes

    Mila · Université de Montréal

  2. Quentin Le Lidec

    New York University

  3. Damien Scieur

    Mila · Université de Montréal · Samsung SAIL

  4. Yann LeCun

    New York University

  5. Randall Balestriero

    Brown University

First published
2026-03-13
Latest paper revision
v3 · 2026-06-03
Venue
arXiv

Covered model versions

Paper ↗PDF ↗

How to access and use

Public code resources
Public access

Code and training / fine-tuning

Access
Public access

Verified on · 2026-11-22

Code license
MIT
Release information

SourceVerified on · 2026-11-22

Checks before deployment

The original release reports simulated visual control and physics evaluation. Performance on physical robots remains unverified in these 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
March 27, 2026
  • Input modalities · Main category · Architecture: arxiv.org
  • Input modalities · Main category · Architecture: github.com

Scope and training

Output

Representations

Derived models

LeWorldModel · SADL shared-BatchNorm ↗Derivative

  • LeWorldModel · SADL shared-BatchNorm: arxiv.org
  • LeWorldModel · SADL shared-BatchNorm: github.com
Submit a source or correction