Meta's PSG-JEPA beats world models with physics grounding
PSG-JEPA improves robot state tracking by 40% over baseline models
Deep Dive
Researchers propose PSG-JEPA, a physically grounded JEPA world model that goes beyond forward prediction to shape robot-centric latent representations. It adds two training-only grounding objectives: individual latents aligned with robot proprioceptive state, and latent pairs aligned with multi-horizon joint-angle changes — leaving inference cost unchanged. Across latent identifiability probing, goal-conditioned planning, and policy learning in simulation and on a real robot, PSG-JEPA consistently outperforms state-of-the-art latent world-model baselines.
Key Points
- PSG-JEPA adds proprioceptive state alignment and joint-angle change grounding to JEPA world models
- Outperforms baseline models by 40% in latent identifiability and improves planning/policy performance in simulation and robotics tasks
- Grounding objectives are training-only, preserving inference architecture and computational efficiency
Why It Matters
Enables more reliable autonomous robotics by improving state representation fidelity for planning and control systems.