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
- Nori Robotics 3D-prints humanoids at $1,688 per robot, targeting 400/month — ycombinator · 2026-09-11
- Ex-NASA Engineer Unveils HAWK, CANARY, PELICAN Drone Fleet to Fight Urban Wildfires — alex_verem · 2026-09-11
- Remote operation of construction machinery from 2,000 km away with under 30ms delay is already real in China — rohanpaul_ai · 2026-09-11
- ESP32 Digital Pets Go on Playdates: Cheap Hardware Meets AI Companions — floguo · 2026-09-11
- TIME P1: What a Year of Living With an Agent-Controlled Desk Device Taught One Team — genmon · 2026-09-11
- Agent with a credit card designs, sources parts and orders a mini DJ controller — bigblueboo · 2026-09-11