Research Project

World Models have become one of the most promising directions for building general, sample-efficient robots: rather than learning a policy that reacts to raw pixels, the agent learns an internal model of how the world evolves and plans inside it. LeRealWorldModel is my open-source take on this idea for real robots. It implements a JEPA (Joint-Embedding Predictive Architecture) latent world model coupled with a GC-IDM amortized planner for goal-conditioned manipulation on the low-cost SO-100 arm, built directly on top of LeRobot and stable-worldmodel.

Instead of predicting future frames pixel by pixel, the model predicts future latent states given actions, which makes learning more stable and inference much faster. Control is closed-loop: a Goal-Conditioned Inverse Dynamics MLP maps (zt, zgoal, horizon) to an action in a single forward pass (~100× faster than sampling-based planning), with a CEM random-shooting planner available as a training-free fallback. The repository ships the full two-stage training pipeline, a pixel-decoder for visualizing what the world model has learned, and identifiability/action-diversity probes. Code is available on GitHub.