Can Predicted Dynamics Exist in the Physical World?

2026-05-23

AI Summary

While existing predictive physics-based AI systems can generate action or state sequences with low RMSE, they often lack physical executability. This work proposes “physical admissibility” as a verification interface between prediction and control, formally encoding physical feasibility through a multi-level validation gate that integrates kinematic constraints, dynamic residuals, and temporal composite conditions, while enabling component-wise violation attribution. Evaluated on the LeRobot PushT task, the method achieves an AUC of 0.957 in identifying infeasible proposals, filtering out 87–89% of invalid candidates while preserving an average task progress of 0.998, thereby substantially enhancing planning reliability.

Abstract

Predictive Physical AI systems output state rollouts, action chunks, and latent plans, yet a low root-mean-square error (RMSE) does not imply that a particular proposal is physically executable. We formulate physical admissibility as a prediction-control interface: before execution, a decoded proposal is treated as candidate dynamics and evaluated using kinematic, dynamic, and direct-to-composed horizon conditions. Passing is not a certificate of task success; rejection identifies violation of the specified physical envelope and gives a component-level reason. On Hugging Face LeRobot PushT, controlled falsification shows that one-step prediction-RMSE and standardized dynamics residuals reach area under the receiver operating characteristic curve (AUC) 0.982 and 0.972, kinematic-only conditions reach AUC 0.592, and the full gate reaches AUC 0.957 with condition-level attribution. In replay-based intervention experiments, residual-based filters and the full physical-admissibility gate prevent 87-$89% of invalid proposals while preserving mean progress near 0.998.

Problem

• physical admissibility
• predictive dynamics
• physical feasibility
• AI-generated proposals
• executable predictions

Innovation

• physical admissibility
• predictive dynamics
• residual-based filtering
• kinematic-dynamic validation
• action proposal verification