lucas-maes/le-wm

Health v2 · maintenance only

Flags: no_releases young

Adoption not part of the score

4344 stars · 634 forks observed · 2026-08-28

What it is AI-extracted, prompt v1, taxonomy v1, 2026-08-29, confidence not recorded

LeWorldModel (LeWM) is the official PyTorch codebase for a JEPA-based world model that trains stably end-to-end from raw pixels using only two loss terms: a next-embedding prediction loss and a Gaussian regularizer. It is a lightweight (~15M parameter) model trainable on a single GPU in hours, supporting planning in 2D and 3D control tasks and probing of latent physical structure.

Use cases

• train a JEPA world model from pixels on a single GPU
• plan in 2D and 3D control environments with a learned latent dynamics model
• reproduce the LeWM paper results from the official codebase
• probe latent representations for encoded physical quantities
• evaluate world models on surprise detection of physically implausible events
• benchmark lightweight world models against foundation-model-based alternatives

When to choose

• you need a small, fast world model trainable on a single GPU
• you want a JEPA that avoids representation collapse without EMA or pretrained encoders
• you are doing research on self-supervised world models or model-based control
• you want to reproduce or extend the LeWM paper

When to avoid

• you need a production-ready robotics or game engine world simulator
• you require large-scale foundation-model-based world models with broad generalization
• you need a no-code or turnkey ML tool rather than a research codebase

Facets

library · maturity active

machine-learning deep-learning simulation llm-training machine-learning reinforcement-learning artificial-intelligence python world-model jepa self-supervised-learning representation-learning model-based-planning research-code pytorch research gpu linux macos

Member repositories

For agents

markdown · JSON · MCP: product_card(name="lucas-maes/le-wm")