SG-JEPA Paper: World Models That Train on Earth and Deploy on Mars, Halving Zero-Shot Physics Error

randall_balestr · x · 2026-09-11

A new arXiv paper (2609.10464) introduces Semigroup-JEPA (SG-JEPA), testing whether JEPA world models can zero-shot generalize to dynamics never seen in training — "train on Earth, deploy on Mars." It conditions the temporal model on physics parameters via action-conditioning and jointly trains encoder and predictor with autoregressive latent rollout. Evaluated on tasks across gravitational fields, SG-JEPA cuts open-loop prediction error up to 2x vs DINO-WM on 2D data and boosts control success up to 2.5x on 3D robotic datasets with diffusion policies. A linear feature model explains the gains: backpropagating multi-step rollout loss into representations trains the encoder to preserve physics-consistent features.

Original post →

More from Embodied

Embodied channel →