jepa-pusht — LeWorldModel (PushT)

Optimize a JEPA world model for the PushT planning task. Minimize validation loss under 16M parameter and 10-minute training time constraints.

Based on LeWorldModel by Lucas Maes et al.

Quickstart

bash prepare.sh
bash eval/eval.sh

Metric

• val_loss (lower is better) — MSE between predicted and target embeddings on validation set.

Constraints

• Max 16M trainable parameters
• Max 10 minutes training time (enforced by Lightning max_time)