Title: Do Predicted Dynamics Have a Place in the Physical Realm?

Abstract: While Predictive Physical AI systems generate state trajectories, action sequences, and latent plans, achieving a low root-mean-square error (RMSE) does not guarantee that a proposed action is physically feasible. This study defines physical admissibility as a prediction-control interface, wherein a decoded proposal is assessed as a candidate dynamic model prior to execution. This evaluation relies on kinematic, dynamic, and direct-to-composed horizon conditions. It is important to note that passing these checks does not certify task success; rather, rejection signals a breach of the defined physical envelope, providing specific reasons at the component level.

In controlled falsification experiments using the Hugging Face LeRobot PushT benchmark, one-step prediction RMSE and standardized dynamics residuals achieved area under the receiver operating characteristic curve (AUC) scores of 0.982 and 0.972, respectively. Kinematic-only conditions yielded an AUC of 0.592, while the comprehensive gate reached an AUC of 0.957, complete with condition-level attribution. Furthermore, replay-based intervention experiments demonstrated that residual-based filters and the full physical-admissibility gate successfully blocked 87–89% of invalid proposals, while maintaining a mean progress metric close to 0.998.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC