Improving Robot World Models via State Grounding

Summary & Methodology Analysis

The paper identifies that standard forward-prediction in Joint-Embedding Predictive Architecture (JEPA) world models fails to enforce reliable identifiability of robot-centric physical states. Because forward-only training leaves an identifiability gap where physical states are not reliably recoverable from latents, downstream planning and policy tasks suffer. Furthermore, inverse-dynamics supervision is often ambiguous, as multiple distinct action sequences can result in the same endpoint state change, making it an unreliable signal for physical transitions. To address this, the authors introduce PSG-JEPA. This approach retains the baseline JEPA encoder and causal action-conditioned predictor architecture, but adds lightweight state and transition heads. The state head maps individual latents to the proprioceptive robot state, while the transition head maps latent pairs to multi-horizon joint-angle changes. These additions act as auxiliary training objectives that force the model to ground its abstract latent representations in measurable physical realities. A key engineering advantage is that both the state and transition grounding heads are discarded after the training process is complete. By removing these heads before deployment, the model maintains its original inference architecture and incurs no additional computational cost for production usage. The authors evaluated this approach against the LeWM baseline on tasks including OGBench-Cube, OGBench-Scene, LIBERO-Goal, and Mobile ALOHA. The results show that PSG-JEPA achieves an average 79.3% success rate compared to 60.0% for the baseline, demonstrating that explicit physical state grounding significantly enhances task performance.