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.
More from Embodied
- Robot learning researcher joins Mila to work on physical AI and RL — pcastr · 2026-10-27
- Full teardown: inside Zmotor's 32mm linear actuator for humanoids — MarwaEldiwiny · 2026-10-27
- Full Teardown of a Zmotor Linear Actuator Used in China's Humanoid Robots — MarwaEldiwiny · 2026-10-27
- World Model Assigns Weight and Friction to Objects, Then Simulates with a Physics Engine — jonstephens85 · 2026-10-27
- First unboxing shots of Apple's foldable "iPhone Duo" surface — chrisfirst · 2026-10-27
- Bedrock Robotics Raises $600M Series C as Autonomous Excavators Do Paid Work on US Job Sites — davidpattersonx · 2026-10-27