PSG-JEPA Grounds Robot World Models with Physical State Awareness
zhenjun_zhao · x · 2026-08-13
To address the lack of reliable physical state identifiability in forward prediction objectives of JEPA world models, researchers proposed PSG-JEPA (Physical State Grounding JEPA).
- The Problem: Current JEPA-based world models learn action-conditioned predictive latent dynamics but fail to explicitly enforce reliable identifiability of robot-centric physical states from individual latents, limiting downstream planning and policy performance.
- The Method: PSG-JEPA introduces two complementary grounding objectives beyond forward prediction: grounding individual latents in robot proprioceptive states, and grounding latent pairs in multi-horizon joint-angle changes. These objectives are applied only during training, leaving inference architecture and computational cost unchanged.
- Results: Across three evaluation levels (latent identifiability via probing, goal-conditioned planning on frozen latents, and policy learning in simulation and on a real robot), PSG-JEPA consistently outperformed state-of-the-art latent world-model baselines.
More from Embodied
- VLAIRobotics Launches Dual-Arm Humanoid K1 Starting at $2,700 — AI寒武纪 · 2026-08-13
- Dyna-2 Pre-Trained on 1M Hours of Human Video Achieves Cross-Embodiment Transfer — iamfakhrealam · 2026-08-13
- Denmark Develops Jointless Earthworm-Inspired Soft Robot for Search and Rescue — lukas_m_ziegler · 2026-08-13
- SHAPER: Train-Free Skill Evolution for Embodied Agents — Peidong Wang · 2026-08-13
- Expert: Humanoid Robots Could Be Trillion-Dollar Market, But Form Factor Depends on Environment — MarwaEldiwiny · 2026-08-13
- Minimax H3 Runs on RTX 5060 Ti: 25-Second Video Takes 93 Minutes to Generate — Beginning_Tip300 · 2026-08-13