lucas-maes

lucas-maes / le-wm

Public

Official code base for LeWorldModel: Stable End-to-End Joint-Embedding Predictive Architecture from Pixels

le-wm.github.io
85
4
100% credibility
Found Mar 23, 2026 at 85 stars -- GitGems finds repos before they trend. Get early access to the next one.
Sign Up Free
AI Analysis
Python
AI Summary

LeWorldModel is a research codebase for training and evaluating compact AI world models that predict future states from raw pixel observations in robotics tasks.

How It Works

1
🔍 Discover LeWorldModel

You stumble upon this cool research project that teaches AI to predict how objects move just by watching pixel videos, perfect for robot brains.

2
💻 Prepare your workspace

You set up a simple virtual space on your computer to get everything ready for training the AI.

3
📥 Download video datasets

You grab ready-made packs of robot action videos from a shared online folder and place them in your storage spot.

4
🧠 Train the AI brain

You start the training process, feeding it videos so it learns to imagine future movements – it hums along for a few hours on your computer's graphics power.

5
🧪 Test on robot tasks

You load your trained AI into test environments like pushing objects or navigating rooms to see how well it plans actions.

🎉 AI masters predictions

Your AI now reliably forecasts movements, detects weird events, and controls robots better and faster than before, ready for real-world experiments.

Sign up to see the full architecture

4 more

Sign Up Free

Star Growth

See how this repo grew from 85 to 85 stars Sign Up Free
Repurpose This Repo

Repurpose is a Pro feature

Generate ready-to-use prompts for X threads, LinkedIn posts, blog posts, YouTube scripts, and more -- with full repo context baked in.

Unlock Repurpose
AI-Generated Review

What is le-wm?

LeWM builds stable world models end-to-end from raw pixels for robotics and control tasks, predicting future states in compact latent spaces without representation collapse. Developers get a PyTorch-based JEPA that trains in hours on a single GPU with ~15M parameters, using simple commands like `python train.py data=pusht` after installing via uv pip and downloading HDF5 datasets. It supports planning in 2D/3D envs like PushT or Cube, with pretrained checkpoints for quick eval via `python eval.py --config-name=pusht.yaml policy=pusht/lewm`.

Why is it gaining traction?

Unlike fragile JEPA alternatives needing multi-term losses or pretraining, LeWM uses just two losses—one for prediction, one for Gaussian regularization—cutting hyperparameters from six to one. Users notice 48x faster planning than foundation models, competitive success rates across tasks, and meaningful latents that probe physical properties or detect anomalies. The official GitHub repo ties to a strong paper with Yann LeCun, plus easy Hydra configs and stable-worldmodel integration.

Who should use this?

RL engineers prototyping pixel-based planners for manipulation (e.g., pushing objects, reaching goals) or navigation (two-room mazes). Robotics devs needing lightweight models for edge deployment, or researchers exploring joint-embedding architectures beyond control, like anomaly detection in sims.

Verdict

Grab it if you're in world modeling—docs are solid, pretrained models work out-of-box, and it delivers on stability claims despite 85 stars and 1.0% credibility score signaling early maturity. Test on your datasets before production.

(198 words)

Sign up to read the full AI review Sign Up Free

Similar repos coming soon.