Semigroup-JEPA: Latent Dynamics Consistency for Zero-Shot Physics Generalization
SG-JEPA’s gains come mainly from learning latent features that survive multi-step physics rollouts.
The paper tests JEPA-style world models on gravity-shifted dynamics, from weak-field floating to strong-field bouncing. SG-JEPA conditions the temporal model on the physics parameter and trains encoder and predictor through autoregressive latent rollout. It reports up to 2x lower open-loop prediction error on 2D datasets and up to 2.5x higher control success on 3D robotic datasets versus DINO-WM. The authors argue the improvement is driven more by the encoder preserving dynamics-relevant features than by a stronger predictor. ArXiv · AI/CL/LG's note
The paper tests JEPA-style world models on gravity-shifted dynamics, from weak-field floating to strong-field bouncing. SG-JEPA conditions the temporal model on the physics parameter and trains encoder and predictor through autoregressive latent rollout. It reports up to 2x lower open-loop prediction error on 2D datasets and up to 2.5x higher control success on 3D robotic datasets versus DINO-WM. The authors argue the improvement is driven more by the encoder preserving dynamics-relevant features than by a stronger predictor. ArXiv · AI/CL/LG's note
score 4